Patent Grant
11719963
This disclosure relates to optical modulation for optoelectronic processing.
Neuromorphic computing is an approach of approximating the operation of a brain in the electronic domain. A prominent approach to neuromorphic computing is an artificial neural network (ANN), which is a collection of artificial neurons that are interconnected in specific ways to process information in a way similar to how a brain functions. ANNs have found uses in a wide range of applications including artificial intelligence, speech recognition, text recognition, natural language processing, and various forms of pattern recognition.
An ANN has an input layer, one or more hidden layers, and an output layer. Each of the layers have nodes, or artificial neurons, and the nodes are interconnected between the layers. Each node of the hidden layers performs a weighted sum of the signals received from nodes of a previous layer, and performs a nonlinear transformation (“activation”) of the weighted sum to generate an output. The weighted sum can be calculated by performing a matrix multiplication step. As such, computing an ANN typically involves multiple matrix multiplication steps, which are typically performed using electronic integrated circuits.
Computation performed on electronic data, encoded in analog or digital form on electrical signals (e.g., voltage or current), is typically implemented using electronic computing hardware, such as analog or digital electronics implemented in integrated circuits (e.g., a processor, application-specific integrated circuit (ASIC), or a system on a chip (SoC)), electronic circuit boards, or other electronic circuitry. Optical signals have been used for transporting data, over long distances, and over shorter distances (e.g., within data centers). Operations performed on such optical signals often take place in the context of optical data transport, such as within devices that are used for switching or filtering optical signals in a network. Use of optical signals in computing platforms has been more limited. Various components and systems for all-optical computing have been proposed. For example, such systems can include conversion from and to electrical signals at the input and output, respectively, but not use both types of signals (electrical and optical) for significant operations that are performed in computations.
In general, in a first aspect, an apparatus includes a substrate on which is disposed an optical waveguide, the optical waveguide including three or more non-overlapping optical waveguide segments located at respective positions along the optical waveguide. The apparatus further includes three or more optical modulators, in which each optical modulator includes a corresponding one of the waveguide segments and is configured to apply an optical modulation that is proportional to a length of that corresponding waveguide segment. The apparatus further includes three or more electrical contacts configured to receive respective bit values of each of a series of binary values, in which each binary value includes at least three bit values including a least significant bit (LSB) bit value, a most significant bit (MSB) bit value, and at least one intermediate bit (IB) bit value between the LSB bit value and the MSB bit value. The apparatus further includes conducting paths configured to electrically couple respective electrical contacts to at least one of the optical modulators, in which each optical modulator is electrically coupled to no more than one of the electrical contacts. The positions of the optical waveguide segments are arranged so that at least one waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving an LSB bit value is positioned between at least one waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving an MSB bit value, and at least one waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving an IB bit value.
The aspects can include one or more of the following features. In some examples, each electrical contact can be electrically coupled to no more than one of the optical modulators.
A relative length of each waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to an LSB bit value of the binary value that includes the particular bit value.
In some examples, at least one electrical contact can be electrically coupled to more than one of the optical modulators.
The lengths of the waveguide segments can be substantially equal to each other, and a number of optical modulators electrically coupled to each electrical contact receiving a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to an LSB bit value of the binary value that includes the particular bit value.
In some examples, a waveguide segment of the optical modulator electrically coupled to the electrical contact receiving an i-th bit value can be positioned between waveguide segments of the optical modulator electrically coupled to the electrical contact configured to receive an (i+1)-th bit value.
In some examples, a waveguide segment of a first optical modulator electrically coupled to a first electrical contact receiving the LSB bit value can be positioned between two waveguide segments of a second optical modulator electrically coupled to a second electrical contact configured to receive a bit value having a bit position that is one bit higher relative to the LSB bit.
Each optical modulator including one of the waveguide segments can include an optical phase modulator.
The optical waveguide segments can be located at respective positions along a portion of the optical waveguide that is between an input coupler and an output coupler of a Mach-Zehnder interferometer.
Each optical modulator including one of the waveguide segments can include an optical amplitude modulator.
An effective refractive index associated with propagation through the optical waveguide can change over a propagation distance through each of the optical waveguide segments before any optical modulation is applied by the optical modulators. A variance of the changing effective refractive index can have a range between a minimum variance and a maximum variance. A portion of the optical waveguide over which the variance is between the minimum variance and an intermediate variance is a low-variance region, in which the intermediate variance is an average of the minimum variance and the maximum variance. A portion of the optical waveguide over which the variance is between the intermediate variance and the maximum variance is a high-variance region. At least one waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving an LSB bit value can be positioned within the low-variance region.
At least one waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving an MSB bit value can overlap with the high-variance region.
In some examples, the effective refractive index can change at least in part due to a change in a fabricated device characteristic.
The fabricated device characteristic can include doping concentration.
In some examples, the effective refractive index can change at least in part due to a change in an operating characteristic.
The operating characteristic can include temperature.
The optical waveguide can include a trim segment having a length that is the same as or less than the waveguide segment of the corresponding optical modulator electrically coupled to the electrical contact receiving the LSB bit value, and the trim segment can be associated with a corresponding optical modulator electrically coupled to an electrical contact receiving a voltage configured to calibrate the apparatus.
The apparatus can include signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the optical modulators in association with a corresponding change between successive digital input values in the series of binary values.
The signal conditioning circuitry can be configured to shape the amplitude change by increasing a size of an amplitude change between a first electrical signal level associated with a first time interval and a second electrical signal level associated with a second time interval for an initial portion of the second time interval.
The signal conditioning circuitry can be configured to shape the amplitude change by decreasing a size of the amplitude change between the first electrical signal level and a second electrical signal level for a final portion of the second time interval.
Each optical waveguide segment can be associated with a diode section positioned along the optical waveguide segment. The diode section can include: a semiconductor diode, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the optical waveguide segment associated with the diode section is modulated in response to a value of a corresponding bit of the digital input value.
The signal conditioning circuitry can be configured to shape the amplitude change by applying an electrical signal to the semiconductor diode through circuitry configured to pump current between the semiconductor diode and a capacitor connected in series between the semiconductor diode and a circuit providing the series of binary values. A quantity of charge transferred by the pumped current can be determined based at least in part on a voltage that is constant over a plurality of successive time intervals in which the series of binary values are provided.
The signal conditioning circuitry can include charge-pump bandwidth enhancement circuits that are configured to drive the optical modulators based on the series of binary values.
At least one optical modulator can include at least one calibration phase shifter that is configured to compensate for an imbalance in optical phase shifts imparted by two waveguide arms of the optical modulator.
In general, in another aspect, a system includes a device configured to perform optical modulations based on an input vector of length N, in which N is an integer greater than 2, and the input vector has N values. The device includes a substrate; a plurality of optical waveguides disposed on the substrate; and a plurality of optical modulators configured to receive electric signals representing the input vector and apply optical modulations to optical signals in the optical waveguides according to the values of the input vector. Each optical modulator includes at least one of the optical waveguides, and each optical modulator is configured to modulate light in the optical waveguide according to one of the values in the input vector. Each optical modulator includes a plurality of sub-modulators, each sub-modulator includes a segment of the waveguide of the optical modulator, and different sub-modulators include different segments of the waveguide. Each sub-modulator is configured to apply the optical modulation on the corresponding segment of the waveguide, and the optical modulation is proportional to a length of the corresponding segment of the waveguide. For each optical modulator, a first one of the sub-modulators is configured to apply the optical modulation based on a least significant bit (LSB) of the digital value, a third one of the sub-modulators is configured to apply the optical modulation based on a most significant bit (MSB) of the digital value, and a second one of the sub-modulators is configured to apply the optical modulations based on an intermediate bit (IB) of the digital value. For each optical modulator, the first sub-modulator is positioned between the second sub-modulator and the third sub-modulator.
The aspects can include one or more of the following features. Each of at least some of the sub-modulators can include at least one of a phase shifter or an optical amplitude modulator.
In some examples, each value of the input vector can include M bits, M being an integer greater than 1, and each bit value can be provided to no more than one sub-modulator.
A relative length of each waveguide segment of a corresponding sub-modulator receiving a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to a least significant bit.
In some examples, each value of the input vector can include M bits, M being an integer greater than 1, and at least one of the bit values can be provided to more than one of the sub-modulators.
The lengths of the waveguide segments of the sub-modulators within an optical modulator can be substantially equal to each other, and a number of sub-modulators configured to receive a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to a least significant bit.
The first segment of the waveguide can have a first level of variance of effective refractive index, the second segment of the waveguide can have a second level of variance of effective refractive index, and the third segment of the waveguide can have a third level of variance of effective refractive index. The first level of variance can be smaller than the second level of variance, and the first level of variance can also be smaller than the third level of variance.
The system can include a processor unit including a light source configured to provide a plurality of light outputs, in which the optical signals in the optical waveguides are derived from the light outputs. The processor unit can further include the plurality of optical modulators, in which the plurality of optical modulators are coupled to the light source.
The system can include signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the optical modulators in association with a corresponding change between successive digital input values.
The signal conditioning circuitry can be configured to shape the amplitude change by increasing a size of an amplitude change between a first electrical signal level associated with a first time interval and a second electrical signal level associated with a second time interval for an initial portion of the second time interval.
The signal conditioning circuitry can be configured to shape the amplitude change by decreasing a size of the amplitude change between the first electrical signal level and a second electrical signal level for a final portion of the second time interval.
Each optical waveguide segment can be associated with a diode section positioned along the optical waveguide segment. The diode section can include: a semiconductor diode, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the optical waveguide segment associated with the diode section is modulated in response to a corresponding bit of the value of the input vector.
The signal conditioning circuitry can be configured to shape the amplitude change by applying an electrical signal to the semiconductor diode through circuitry configured to pump current between the semiconductor diode and a capacitor connected in series between the semiconductor diode and a circuit providing the values of the input vector. A quantity of charge transferred by the pumped current can be determined based at least in part on a voltage that is constant over a plurality of successive time intervals in which the values of the input vector are provided.
The signal conditioning circuitry can include charge-pump bandwidth enhancement circuits that are configured to drive the optical modulators based on the electric signals representing the input vector.
At least one optical modulator can include at least one calibration phase shifter that is configured to compensate for an imbalance in optical phase shifts imparted by two waveguide arms of the optical modulator.
In general, in another aspect, a system includes a first unit configured to generate a plurality of modulator control signals; and a processor unit. The processor unit includes a light source configured to provide a plurality of light outputs; a plurality of optical modulators coupled to the light source and the first unit, and a matrix multiplication unit. The plurality of optical modulators are configured to generate an optical input vector by modulating the plurality of light outputs provided by the light source based on the plurality of modulator control signals, the optical input vector including a plurality of optical signals. Each of the optical modulators has a segmented design. The matrix multiplication unit is coupled to the plurality of optical modulators and the first unit, in which the matrix multiplication unit is configured to transform the optical input vector into an analog output vector based on a plurality of weight control signals. The system further includes a second unit coupled to the matrix multiplication unit and configured to convert the analog output vector into a digitized output vector. The system further includes a controller including integrated circuitry configured to perform operations. The operations include receiving an artificial neural network computation request including an input dataset that includes a first digital input vector; and receiving a first plurality of neural network weights. The operations further include generating, through the first unit, the first plurality of modulator control signals and the first plurality of weight control signals. The first plurality of modulator control signals are generated based on the first digital input vector, and the first plurality of weight control signals are generated based on the first plurality of neural network weights.
The aspects can include one or more of the following features. Each modulator having the segmented design can include an optical waveguide including three or more non-overlapping optical waveguide segments located at respective positions along the optical waveguide, in which each waveguide segment is part of a sub-modulator configured to apply an optical modulation that is proportional to a length of that waveguide segment. Each modulator can further include three or more electrical contacts configured to receive respective bit values of each of a series of binary values, in which each binary value includes at least three bit values including a least significant bit (LSB) bit value, a most significant bit (MSB) bit value, and at least one intermediate bit (IB) bit value between the LSB bit value and the MSB bit value. Each modulator can further include conducting paths configured to electrically couple respective electrical contacts to at least one of the sub-modulators, in which each sub-modulator is electrically coupled to no more than one of the electrical contacts. The positions of the optical waveguide segments can be arranged so that at least one waveguide segment of a corresponding sub-modulator electrically coupled to an electrical contact receiving an LSB bit value is positioned between at least one waveguide segment of a corresponding sub-modulator electrically coupled to an electrical contact receiving an MSB bit value, and at least one waveguide segment of a corresponding sub-modulator electrically coupled to an electrical contact receiving an IB bit value.
In some examples, each electrical contact can be electrically coupled to no more than one of the sub-modulators.
A relative length of each waveguide segment of a corresponding optical modulator electrically coupled to an electrical contact receiving a particular bit value can be proportional to 2n, where n is a bit position of the particular bit value relative to an LSB bit value of the binary value that includes the particular bit value.
In some examples, at least one electrical contact is electrically coupled to more than one of the sub-modulators.
The lengths of the waveguide segments can be substantially equal to each other, and a number of sub-modulators electrically coupled to each electrical contact receiving a particular bit value can be proportional to 2n, where n is a bit position of the particular bit value relative to an LSB bit value of the binary value that includes the particular bit value.
A waveguide segment of the sub-modulator electrically coupled to the electrical contact receiving an i-th bit value can be positioned between waveguide segments of the sub-modulator electrically coupled to the electrical contact configured to receive an (i+1)-th bit value.
A waveguide segment of a first sub-modulator electrically coupled to a first electrical contact receiving the LSB bit value can be positioned between two waveguide segments of a second sub-modulator electrically coupled to a second electrical contact configured to receive a bit value having a bit position that is one bit higher relative to the LSB bit.
Each sub-modulator including one of the waveguide segments can include an optical phase modulator.
The optical waveguide segments can be located at respective positions along a portion of the optical waveguide that is between an input coupler and an output coupler of a Mach-Zehnder interferometer.
Each sub-modulator can including one of the waveguide segments can include an optical amplitude modulator.
An effective refractive index associated with propagation through the optical waveguide can change over a propagation distance through each of the optical waveguide segments before any optical modulation is applied by the optical modulators. A variance of the changing effective refractive index can have a range between a minimum variance and a maximum variance. A portion of the optical waveguide over which the variance is between the minimum variance and an intermediate variance can be a low-variance region, in which the intermediate variance can be an average of the minimum variance and the maximum variance. A portion of the optical waveguide over which the variance is between the intermediate variance and the maximum variance can be a high-variance region. At least one waveguide segment of a corresponding sub-modulator electrically coupled to an electrical contact receiving an LSB bit value can be positioned within the low-variance region.
At least one waveguide segment of a corresponding sub-modulator electrically coupled to an electrical contact receiving an MSB bit value can overlap with the high-variance region.
The effective refractive index can change at least in part due to a change in a fabricated device characteristic.
The fabricated device characteristic can include doping concentration.
The effective refractive index can change at least in part due to a change in an operating characteristic.
The operating characteristic can include temperature.
The optical waveguide can include a trim segment having a length that is the same as or less than the waveguide segment of the corresponding sub-modulator electrically coupled to the electrical contact receiving the LSB bit value, and the trim segment can be associated with a corresponding optical modulator electrically coupled to an electrical contact receiving a voltage configured to calibrate the apparatus.
The system can include signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the sub-modulators in association with a corresponding change between successive digital input values in the series of binary values.
The signal conditioning circuitry can be configured to shape the amplitude change by increasing a size of an amplitude change between a first electrical signal level associated with a first time interval and a second electrical signal level associated with a second time interval for an initial portion of the second time interval.
The signal conditioning circuitry can be configured to shape the amplitude change by decreasing a size of the amplitude change between the first electrical signal level and a second electrical signal level for a final portion of the second time interval.
Each optical waveguide segment can be associated with a diode section positioned along the optical waveguide segment. The diode section can include: a semiconductor diode, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the optical waveguide segment associated with the diode section is modulated in response to a value of a corresponding bit of the digital input value.
The signal conditioning circuitry can be configured to shape the amplitude change by applying an electrical signal to the semiconductor diode through circuitry configured to pump current between the semiconductor diode and a capacitor connected in series between the semiconductor diode and a circuit providing the series of binary values. A quantity of charge transferred by the pumped current can be determined based at least in part on a voltage that is constant over a plurality of successive time intervals in which the series of binary values are provided.
The signal conditioning circuitry can include charge-pump bandwidth enhancement circuits that are configured to drive the optical modulators based on the modulator control signals.
At least one optical modulator can include at least one calibration phase shifter that is configured to compensate for an imbalance in optical phase shifts imparted by two waveguide arms of the optical modulator.
In general, in another aspect, a method for controlling optical modulation includes receiving input including a series of binary values, in which each binary value includes at least three bit values including a least significant bit (LSB) bit value, a most significant bit (MSB) bit value, and at least one intermediate bit (IB) bit value between the LSB bit value and the MSB bit value. The method further includes receiving an optical wave that propagates along an optical waveguide that includes three or more non-overlapping optical waveguide segments located at respective positions along the optical waveguide, in which each waveguide segment is associated with a corresponding optical modulator configured to apply an optical modulation that is proportional to a length of that waveguide segment. The method further includes providing respective bit values of a first binary value of the series of binary values to at least one of the optical modulators, in which each optical modulator is configured to receive no more than one of the bit values of the first binary value. The positions of the optical waveguide segments are arranged so that at least one waveguide segment associated with a corresponding optical modulator configured to receive an LSB bit value is positioned between at least one waveguide segment associated with a corresponding optical modulator configured to receive an MSB bit value, and at least one waveguide segment associated with a corresponding optical modulator configured to receive an IB bit value.
The aspects can include one or more of the following features. In some examples, each bit value of the first binary value can be received by no more than one of the optical modulators.
A relative length of each waveguide segment associated with a corresponding optical modulator configured to receive a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to an LSB bit value of the first binary value.
In some examples, at least one bit value of the first binary value can be received by more than one of the optical modulators.
The lengths of the waveguide segments can be substantially equal to each other, and a number of optical modulators configured to receive a particular bit value can be proportional to 2n, in which n is a bit position of the particular bit value relative to an LSB bit value of the first binary value.
Each optical modulator associated with one of the waveguide segments can include an optical phase modulator.
The optical waveguide segments can be located at respective positions along a portion of the optical waveguide that is between an input coupler and an output coupler of a Mach-Zehnder interferometer.
Each optical modulator associated with one of the waveguide segments can include an optical amplitude modulator.
An effective refractive index associated with propagation through the optical waveguide can change over a propagation distance through each of the optical waveguide segments before any optical modulation is applied by the optical modulators. A variance of the changing effective refractive index can have a range between a minimum variance and a maximum variance. A portion of the optical waveguide over which the variance is between the minimum variance and an intermediate variance is a low-variance region, in which the intermediate variance is an average of the minimum variance and the maximum variance. A portion of the optical waveguide over which the variance is between the intermediate variance and the maximum variance is a high-variance region. At least one waveguide segment associated with a corresponding optical modulator configured to receive an LSB bit value can be positioned within the low-variance region.
At least one waveguide segment associated with a corresponding optical modulator configured to receive an MSB bit value can overlap with the high-variance region.
In some examples, the effective refractive index can change at least in part due to a change in a fabricated device characteristic.
The fabricated device characteristic can include doping concentration.
In some examples, the effective refractive index can change at least in part due to a change in an operating characteristic.
The operating characteristic can include temperature.
The optical waveguide can include a trim segment having a length that is the same as or less than the waveguide segment of the corresponding optical modulator electrically coupled to the electrical contact receiving the LSB bit value, and the trim segment can be associated with a corresponding optical modulator electrically coupled to an electrical contact. The method can include providing a voltage signal to the electrical contract electrically coupled to the optical modulator associated with the trim segment to calibrate the optical modulation of the optical wave.
The method can include shaping an amplitude change in an electrical signal applied to at least one of the optical modulators in association with a corresponding change between successive digital input values in the series of binary values by increasing a size of an amplitude change between a first electrical signal level associated with a first time interval and a second electrical signal level associated with a second time interval for an initial portion of the second time interval.
The method can include shaping the amplitude change by decreasing a size of the amplitude change between the first electrical signal level and a second electrical signal level for a final portion of the second time interval.
Each optical waveguide segment can be associated with a diode section positioned along the optical waveguide segment, and the diode section can include a semiconductor diode. The method can include applying an electrical signal representing a bit value of the input binary values to the corresponding semiconductor diode in a forward-biased state and modulating an optical property of the optical waveguide segment associated with the diode section.
The method can include shaping an amplitude change in the electrical signal applied to the semiconductor diode by pumping current to a capacitor connected in series with the semiconductor diode. A quantity of charge transferred by pumping current can be determined based at least in part on a voltage that is constant over a plurality of successive time intervals in which the series of binary values are provided.
The optical modulators can include MZI modulators, and the method can include calibrating each of the MZI modulators using calibration phase shifters.
In general, in another aspect, a system includes at least one input optical waveguide configured to receive an optical wave; at least one digital input port configured to receive a series of digital input values over successive time intervals, each digital input value comprising two or more bits; and an optical modulator coupled to the input optical waveguide. The optical modulator includes an optical waveguide portion including multiple diode sections positioned along the optical waveguide portion. The diode sections apply different respective modulation contributions to an optical wave propagating through the optical waveguide portion, and each respective diode section includes a semiconductor diode that has an optical path length of less than about one millimeter, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the diode section is modulated in response to a value of a corresponding bit of the digital input value. The optical modulator includes signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the semiconductor diodes in association with a corresponding change between successive digital input values in the series of digital input values.
Aspects can include one or more of the following features. Shaping the amplitude changes in electrical signals applied to the semiconductor diodes in association with successive digital input values in the series of digital input values can include: increasing a size of an amplitude change between a first electrical signal level associated with a first time interval and a second electrical signal level associated with a second time interval for an initial portion of the second time interval.
Shaping the amplitude changes in electrical signals applied to the semiconductor diodes in association with successive digital input values in the series of digital input values can further include: decreasing a size of the amplitude change between the first electrical signal level and a second electrical signal level for a final portion of the second time interval.
Shaping the amplitude changes in electrical signals applied to the semiconductor diodes in association with successive digital input values in the series of digital input values can include: applying an electrical signal to the semiconductor diode through a matching electrical circuit that is configured to match an impedance associated with the semiconductor diode without significantly changing an amplitude of the applied electrical signal.
The matching electrical circuit can include a passive electrical circuit.
The matching electrical circuit can consists essentially of an inductor.
Shaping the amplitude changes in electrical signals applied to the semiconductor diodes for successive digital input values in the series of digital input values can include: applying an electrical signal to the semiconductor diode through circuitry configured to pump current between the semiconductor diode and a capacitor connected in series between the semiconductor diode and a circuit providing the series of digital input values, in which a quantity of charge transferred by the pumped current is determined based at least in part on a voltage that is constant over a plurality of successive time intervals in which the series of digital input values are provided.
The optical modulator can include an interferometric optical modulator that further includes an optical interference portion configured to provide a level of destructive optical interference that results in a predetermined amplitude reduction based on the cumulative modulation contributions of the diode sections.
The optical interference portion can include an optical combiner.
The optical waveguide portion can include: at least two optical waveguide segments that each receives an optical wave split from the same optical splitter coupled to the input optical waveguide, and provides an optical wave to the optical combiner.
The optical property of the diode section can include an effective refractive index of the diode section, and the different respective modulation contributions can include different respective phase shifts.
An optical path length of a first semiconductor diode of a first diode section can be about twice an optical path length of a second semiconductor diode of a second diode section.
Shaping the amplitude changes in electrical signals applied to the semiconductor diodes for successive digital input values in the series of digital input values can include: applying, to electrical contacts of the first diode section and to electrical contacts of the second diode section, a first predetermined shape of an electrical signal amplitude in response to a change in a value of a corresponding bit from zero to one, and a second predetermined shape of an electrical signal amplitude in response to a change in a value of a corresponding bit from one to zero.
Shaping the amplitude changes in electrical signals applied to the semiconductor diodes for successive digital input values in the series of digital input values can include: applying an electrical signal to electrical contacts of the first diode section through a first matching electrical circuit that is configured to match an impedance associated with the first semiconductor diode, and applying an electrical signal to electrical contacts of the second diode section through a second matching electrical circuit that is configured to match an impedance associated with the second semiconductor diode.
The first matching electrical circuit and the second matching electrical circuit can each consist essentially of an inductor.
The input optical waveguide can be coupled to an optical demultiplexer that separates at least two optical waves of different wavelengths.
The optical modulator can include an absorption optical modulator configured to provide a level of absorbance that results in a predetermined amplitude reduction based on the cumulative modulation contributions of the diode sections.
The optical property of the diode section can include an absorption coefficient of the diode section, and the different respective modulation contributions can include different respective absorbances.
The at least one input optical waveguide can include a plurality of input optical waveguides, a plurality of optical modulators can each be coupled to different respective input optical waveguides of the plurality of input optical waveguides, and outputs from the plurality of optical modulators can be combined to provide a result of vector-matrix multiplication.
In another general aspect, a system includes: at least one input optical waveguide configured to receive an optical wave; at least one digital input port configured to receive a series of digital input values over successive time intervals, each digital input value comprising two or more bits; and an interferometric optical modulator coupled to the input optical waveguide. The interferometric optical modulator includes an optical waveguide portion including a diode section along the optical waveguide portion. The diode section modulates an optical wave propagating through the optical waveguide portion, and the diode section includes: a semiconductor diode that has an optical path length of less than about one millimeter, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the diode section is modulated in response to the digital input value. The interferometric optical modulator includes signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to the semiconductor diode in association with a corresponding change between successive digital input values in the series of digital input values. The signal conditioning circuitry includes: a first signal conditioning path that provides an unconditioned electrical signal corresponding to the series of digital input values, a second signal conditioning path that provides a delayed, scaled, and/or inverted version of the unconditioned electrical signal, and a third signal conditioning path that provides a delayed, scaled, and/or inverted version of the unconditioned electrical signal. The interferometric optical modulator further includes an optical interference portion configured to provide a level of destructive optical interference that results in a predetermined amplitude reduction based on the modulation of the optical property of the diode section.
Aspects can include one or more of the following features. The optical interference portion can include coupling portions of each optical waveguide of a pair of optical waveguides, in which the coupling portions of each optical waveguide are in proximity to each other.
The optical waveguide portion can include a first optical waveguide of the pair of optical waveguides that is formed in a closed path.
In another general aspect, a system includes at least one input optical waveguide configured to receive an optical wave; at least one digital input port configured to receive a series of digital input values, each digital input value including two or more bits; and an optical modulator coupled to the input optical waveguide, the optical modulator including an optical waveguide portion that includes multiple diode sections positioned along the optical waveguide portion, in which the diode sections are configured to apply different respective modulation contributions to an optical wave propagating through the optical waveguide portion.
Aspects can include one or more of the following features. The optical modulator can include signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the diode sections in association with a corresponding change between successive digital input values in the series of digital input values.
Each diode section can include: a semiconductor diode, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the diode section is modulated in response to a value of a corresponding bit of the digital input value.
Each semiconductor diode can have an optical path length of less than about one millimeter.
In another general aspect, a system includes: at least one input optical waveguide configured to receive an optical wave; at least one digital input port configured to receive a series of digital input values, each digital input value including two or more bits; and an interferometric optical modulator coupled to the at least one input optical waveguide. The interferometric optical modulator includes an optical waveguide portion including a diode section positioned along the optical waveguide portion, in which the diode section is configured to modulate an optical wave propagating through the optical waveguide portion. The interferometric optical modulator includes signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to the diode section in association with a corresponding change between successive digital input values in the series of digital input values. The signal conditioning circuitry includes: a first signal conditioning path that provides an unconditioned electrical signal corresponding to the series of digital input values, a second signal conditioning path that provides a delayed, scaled, and/or inverted version of the unconditioned electrical signal, and a third signal conditioning path that provides a delayed, scaled, and/or inverted version of the unconditioned electrical signal.
Aspects can include one or more of the following features. The diode section can include a semiconductor diode that has an optical path length of less than about one millimeter.
The interferometric optical modulator can include an optical interference portion configured to provide a level of destructive optical interference that results in a predetermined amplitude reduction based on the modulation of the optical property of the diode section.
The diode section can include: a semiconductor diode, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the diode section is modulated in response to the digital input value.
The semiconductor diode can have an optical path length of less than about one millimeter.
In another general aspect, a system includes an optical modulator configured to modulate an optical wave representing a series of digital input values, each digital input value including two or more bits. The optical modulator includes an optical waveguide portion that includes multiple diode sections positioned along the optical waveguide portion. The diode sections are configured to apply different respective modulation contributions to the optical wave as the optical wave propagates through the optical waveguide portion. Each diode section includes a semiconductor diode that is configured to operate in a forward-biased state in which an optical property of the diode section is modulated in response to a value of a corresponding bit of the digital input value. The optical modulator includes signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the diode sections in association with a corresponding change between successive digital input values in the series of digital input values.
Aspects can include the following feature. Each semiconductor diode can have an optical path length of less than about one millimeter.
In another general aspect, a system includes: a first unit configured to generate a first set of modulator control signals, a second unit configured to generate a second set of modulator control signals, and a processor unit. The processor unit includes a light source or port configured to provide a plurality of light outputs, and a first set of optical modulators coupled to the light source or port and the first unit. The optical modulators in the first set are configured to generate an optical input vector by modulating the plurality of light outputs provided by the light source or port based on digital input values corresponding to the first set of modulator control signals, the optical input vector comprising a plurality of optical signals. The processor unit also includes a matrix multiplication unit that includes a second set of optical modulators. The matrix multiplication unit is coupled to the second unit, and is configured to transform the optical input vector into an analog output vector based on a plurality of digital weight values corresponding to the second set of modulator control signals. At least one optical modulator in the first set of optical modulators includes a first optical waveguide portion that includes multiple diode sections positioned along the first optical waveguide portion. The diode sections are configured to apply different respective modulation contributions to an optical wave propagating through the first optical waveguide portion based on different respective single bits of one of the digital input values. At least one optical modulator in the second set of optical modulators includes a second optical waveguide portion that includes a single diode section configured to apply a modulation to an optical wave propagating through the second optical waveguide portion based on multiple bits of one of the digital weight values.
Aspects can include the following feature. Each respective diode section includes a semiconductor diode configured to operate in a forward-biased state in which an optical property of the diode section is modulated in response to a value of a corresponding bit of one of the digital input values.
In another general aspect, a system includes: at least one input optical waveguide configured to receive an optical wave; at least one digital input port configured to receive a series of digital input values over successive time intervals, each digital input value including two or more bits; and an optical modulator coupled to the input optical waveguide. The optical modulator includes: an optical waveguide portion including multiple optical waveguide segments that are associated with multiple diode sections positioned along the corresponding optical waveguide segment, the optical waveguide segments being part of a continuous optical waveguide, in which the diode sections are configured to apply different respective modulation contributions to an optical wave propagating through the optical waveguide segments. Each respective diode section includes: a semiconductor diode that is associated with an optical path length of less than about one millimeter, and electrical contacts for applying an electrical signal to the semiconductor diode in a forward-biased state in which an optical property of the optical waveguide segment associated with the diode section is modulated in response to a value of a corresponding bit of the digital input value, and signal conditioning circuitry configured to shape an amplitude change in an electrical signal applied to at least one of the semiconductor diodes in association with a corresponding change between successive digital input values in the series of digital input values.
In another general aspect, a system includes: at least one input optical waveguide configured to receive an optical wave; at least one input port configured to receive a series of modulator control signals; and an optical modulator coupled to the input optical waveguide configured to provide pulse amplitude modulation with four or more amplitude levels. The optical modulator includes a first modulator arm and a second modulator arm, at least one of the first or second modulator arm includes an optical waveguide and multiple phase shifters positioned along the optical waveguide, the phase shifters are configured to apply different respective modulation contributions to an optical wave propagating through the optical waveguide, each phase shifter is coupled to a respective signal conditioning circuitry configured to provide enhanced bandwidth for binary modulation, and different phase shifters are coupled to different signal conditioning circuitry. Each phase shifter includes a semiconductor diode or a capacitor that is associated with an optical path length of less than about one millimeter. The system includes control circuitry configured to, for each phase shifter, at least one of (i) provide an electrical signal to the semiconductor diode to cause the semiconductor diode to be forward-biased in which an optical property of the optical waveguide associated with the phase shifter is modulated in response to the modulator control signals, or (ii) provide an electrical signal to the capacitor to cause electric charges to accumulate at the capacitor in which an optical property of the optical waveguide associated with the phase shifter is modulated in response to the modulator control signals.
In another general aspect, a system includes: an optical modulator configured to modulate an optical wave representing a series of digital input values, each digital input value comprising two or more bits, the optical modulator comprising an optical waveguide portion that includes multiple diode sections positioned along the optical waveguide portion, in which the diode sections are configured to apply different respective modulation contributions to the optical wave as the optical wave propagates through the optical waveguide portion. Each diode section includes a semiconductor diode that is configured to operate in a forward-biased state in which an optical property of the diode section is modulated in response to a value of a corresponding bit of the digital input value. The optical modulator includes a plurality of signal conditioning circuits configured to shape an amplitude change in an electrical signal applied to each of the diode sections in association with a corresponding change between successive digital input values in the series of digital input values, in which each signal conditioning circuit is associated with one of the diode sections, and different diode sections are associated with different signal conditioning circuits.
In another general aspect, a system includes: a first set of optical modulators configured to generate an optical input vector by modulating a plurality of input light signals based on modulator control signals, the optical input vector comprising a plurality of optical signals; and a matrix processing unit comprising a second set of optical modulators, the matrix processing unit is coupled to the modulator array and is configured to transform the optical input vector into an analog output vector based on a plurality of weight values. At least one optical modulator in the first set of optical modulators includes a first modulator arm and a second modulator arm, at least one of the first or second modulator arm includes an optical waveguide and at least two phase shifters positioned along the optical waveguide, each phase shifter is configured to apply a modulation contribution to an optical wave propagating through the optical waveguide based on one of the modulation control signals, different phase shifters are configured to apply different respective modulation contributions to the optical wave propagating through the optical waveguide in response to the same modulation control signal level. At least one optical modulator in the second set of optical modulators includes a first modulator arm and a second modulator arm, each of the first and second modulator arms includes an optical waveguide, at least one of the first and second modulator arms includes a single phase shifter positioned along the optical waveguide, each of the first and second modulator arms includes no more than one phase shifter.
In another general aspect, a system includes: a first unit configured to generate a plurality of modulator control signals; and a processor unit. The processor unit includes: a light source configured to provide a plurality of light outputs, a modulator array, and a matrix processing unit. The modulator array includes a first set of plurality of optical modulators coupled to the light source and the first unit, the plurality of optical modulators being configured to generate an optical input vector by modulating the plurality of light outputs provided by the light source based on the plurality of modulator control signals, the optical input vector including a plurality of optical signals, in which each of the plurality of optical modulators has a segmented design and includes two or more phase shifters, and each of the phase shifters is associated with a signal conditioning circuit configured to implement pre-emphasis and de-emphasis of the phase shifter to enhance an operating bandwidth of the phase shifter. The matrix processing unit is coupled to the modulator array and the first unit, the matrix processing unit being configured to transform the optical input vector into an analog output vector based on a plurality of weight control signals, in which the matrix processing unit includes a second set of plurality of optical modulators that have a non-segmented design, each optical modulator has a single phase shifter in at least one modulator arm, each modulator arm has no more than one phase shifter. The system includes a second unit coupled to the matrix multiplication unit and configured to convert the analog output vector into a digitized output vector; and a controller including integrated circuitry configured to perform operations including: receiving an artificial neural network computation request including an input dataset that includes a first digital input vector; receiving a first plurality of neural network weights; and generating, through the first unit, a first plurality of modulator control signals based on the first digital input vector and a first plurality of weight control signals based on the first plurality of neural network weights.
In another general aspect, an ANN computing method includes: receiving a digital electronic input vector; converting the digital electronic input vector to an analog optical input vector using a modulator array comprising a first set of optical modulators using a segmented design; processing the analog optical input vector using a matrix processing unit that comprises a second set of optical modulators using a non-segmented design; and driving the first set of optical modulators using charge-pump bandwidth enhancement circuits based on the digital electronic input vector or a processed version of the digital electronic input vector. The first set of optical modulators are reconfigured at a first frequency, the second set of optical modulators are reconfigured at a second frequency, and the first frequency is greater than the second frequency.
In another general aspect, an ANN computing method includes: receiving a digital electronic input value having at least a first bit and a second bit; generating a first modulator control signal based on the first bit of the digital electronic input value; generating a second modulator control signal based on the second bit of the digital electronic input value; controlling a first phase shifter of an optical modulator based on the first modulator control signal to cause the first phase shifter to impart a first amount of phase shift to an optical wave propagating in a first waveguide; and controlling a second phase shifter of the optical modulator based on the second modulator control signal to cause the second phase shifter to impart a second amount of phase shift to the optical wave propagating in the first waveguide. The second amount of phase shift imparted by the second phase shifter is weighted twice the first amount of phase shift imparted by the first phase shifter. The method includes driving the first phase shifter using a first charge-pump bandwidth enhancement circuit based on the first modulator control signal; and driving the second phase shifter using a second charge-pump bandwidth enhancement circuit based on the first modulator control signal.
Aspects can have one or more of the following advantages.
In some embodiments described herein, the type of components that are used to perform optical modulation, and/or the characteristics of those components, are selected to provide performance improvements or other design advantages in the resulting optoelectronic computing system. In a system that is fabricated using a silicon photonics technology, an optical waveguide can be formed in the silicon and there are various types of semiconductor diode structures that can be formed by doping the silicon in proximity to the waveguide to enable modulation of an optical wave propagating in the waveguide. For example, PIN diode structures or metal-oxide-semiconductor (MOS) capacitors can be designed to enable free-carrier-based modulation by carrier injection, carrier depletion, or carrier accumulation, using different doping profiles. Carrier injection uses a forward-biased PIN diode structure that typically has a relatively small size (e.g., shorter than 1 mm) due to the high modulation efficiency but provides lower-speed modulation (e.g., less than around 1 Gb/s), and carrier depletion uses a reverse-biased PIN diode structure that typically has a larger size (e.g., longer than 1 mm) but provides higher-speed modulation (e.g., greater than around 1 Gb/s). In some examples, carrier accumulation uses a capacitor to store charges in which the capacitor includes a thin vertical insulation layer. The thin vertical insulation layer can be difficult to fabricate for some semiconductor processes. Thus, for some semiconductor manufacturing processes, the forward-biased PIN diode structure for implementing carrier injection can be easier to fabricate.
Carrier injection provides both small size and low power characteristics that are useful in large scale optoelectronic computing system that has a dense array of modulators. Using a segmented modulator design for multi-bit modulation (e.g., PAM modulation) in combination with a carrier injection PIN diode structure, as described in more detail below, it is possible to use bandwidth-enhancement techniques to overcome the bandwidth limitations (e.g., increasing the bandwidth by around a factor of 10) in a simple and compact design that is suitable for the dense arrangement of modulators within the system. Since the resulting enhanced bandwidth does not necessarily need to be as high as can be achieved in a carrier depletion or carrier accumulation modulator, the combination of simple bandwidth enhancement enabled by the segmented modulator and the compact size of a carrier injection modulator synergistically provide advantages that are especially useful in some of the system embodiments described herein. Some forms of bandwidth enhancement, such as pre-emphasis, can be challenging to implement when there are more than two amplitude levels. But, with a segmented modulator, individual binary pre-emphasis can be implemented for the two amplitude levels of each bit of a multi-bit modulation, as described in more detail below.
Additionally, because of the computations being performed using the modulated optical signals, the number of analog amplitude levels of the results of those computations being sampled when converting to digital signals (e.g., 8-bit signals using PAM-256 modulation) can be larger than the number of levels of the inputs to those computations (e.g., 4-bit signals using PAM-16 modulation). For example, a 4-bit value multiplied by another 4-bit value can result in a value that is suitably represented by 8 bits. Thus, the use of bandwidth-enhancement to increase the signal to noise ratio is also motivated by the need for effective analog-to-digital conversion in the context of the increase in resolution needed after the input signals are processed by the optoelectronic computing system.
Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. In case of conflict with patent applications or patent application publications incorporated herein by reference, the present specification, including definitions, will control.
The details of one or more embodiments of the subject matter described in this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the invention will become apparent from the description, the drawings, and the claims.
The disclosure is best understood from the following detailed description when read in conjunction with the accompanying drawings. It is emphasized that, according to common practice, the various features of the drawings are not to-scale. On the contrary, the dimensions of the various features are arbitrarily expanded or reduced for clarity.
Like reference numbers and designations in the various drawings indicate like elements.
The integrated circuitry of the controller 110 can be an application specific integrated circuit specifically configured to perform the steps of an ANN computation process. For example, the integrated circuitry can implement a microcode or a firmware specific to performing the ANN computation process. As such, the controller 110 can have a reduced set of instructions relative to a general purpose processor used in conventional computers, such as the computer 102. In some implementations, the integrated circuitry of the controller 110 can include two or more circuitries configured to perform different steps of the ANN computation process.
In an example operation of the ANN computation system 100, the computer 102 can issue an artificial neural network computation request to the ANN computation system 100. The ANN computation request can include neural network weights that define an ANN, and an input dataset to be processed by the provided ANN. The controller 110 receives the ANN computation request, and stores the input dataset and the neural network weights in the memory unit 120.
The input dataset can correspond to various digital information to be processed by the ANN. Examples of the input dataset include image files, audio files, LiDAR point cloud, biological data files, and GPS coordinates sequences, and the operation of the ANN computation system 100 will be described based on receiving an image file as the input dataset. In general, the size of the input dataset can vary greatly, from hundreds of data points to millions of data points or larger. For example, a digital image file with a resolution of 1 megapixel has approximately one million pixels, and each of the one million pixels can be a data point to be processed by the ANN. Due to the large number of data points in a typical input dataset, the input dataset is typically divided into multiple digital input vectors of smaller size to be individually processed by the optoelectronic processor 140. As an example, for a greyscale digital image, the elements of the digital input vectors can be 8-bit values representing the intensity of the image, and the digital input vectors can have a length that ranges from 10's of elements (e.g., 32 elements, 64 elements) to hundreds of elements (e.g., 256 elements, 512 elements). In general, input dataset of arbitrary size can be divided into digital input vectors of a size suitable for processing by the optoelectronic processor 140. In cases where the number of elements of the input dataset is not divisible by the length of the digital input vector, zero padding can be used to fill out the data set to be divisible by the length of the digital input vector. The processed outputs of the individual digital input vectors can be processed to reconstruct a complete output that is a result of processing the input dataset through the ANN. In some implementations, the dividing of the input data set into multiple input vectors and subsequent vector-level processing can be implemented using block matrix multiplication techniques.
The neural network weights are a set of values that define the connectivity of the artificial neurons of the ANN, including the relative importance, or weights, of those connections. An ANN can include one or more hidden layers with respective sets of nodes. In the case of an ANN with a single hidden layer, the ANN can be defined by two sets of neural network weights, one set corresponding to the connectivity between the input nodes and the nodes of the hidden layer, and a second set corresponding to the connectivity between the hidden layer and the output nodes. Each set of neural network weights that describes the connectivity corresponds to a matrix to be implemented by the optoelectronic processor 140. For ANNs with two or more hidden layers, additional sets of neural network weights are needed to define the connectivity between the additional hidden layers. As such, in general, the neural network weights included in the ANN computation request can include multiple sets of neural network weights that represent the connectivity between various layers of the ANN.
As the input dataset to be processed is typically divided into multiple smaller digital input vectors for individual processing, the input dataset is typically stored in a digital memory. However, the speed of memory operations between a memory and a processor of the computer 102 can be significantly slower than the rate at which the ANN computation system 100 can perform ANN computations. For example, the ANN computation system 100 can perform tens to hundreds of ANN computations during a typical memory read cycle of the computer 102. As such, the rate at which ANN computations can be performed by the ANN computation system 100 can be limited below its full processing rate if an ANN computation by the ANN computation system 100 involves multiple data transfers between the system 100 and the computer 102 during the course of processing an ANN computation request. For example, if the computer 102 were to access the input dataset from its own memory and provide the digital input vectors to the controller 110 when requested, the operation of the ANN computation system 100 would likely be greatly slowed down by the time needed for the series of data transfers that would be needed between the computer 102 and the controller 110. It should be noted that a memory access latency of the computer 102 is typically non-deterministic, which further complicates and degrades the speed at which digital input vectors can be provided to the ANN computation system 100. Further, the processor cycles of the computer 102 can be wasted on managing the data transfer between the computer 102 and the ANN computation system 100.
Instead, in some implementations, the ANN computation system 100 stores the entire input dataset in the memory unit 120, which is a part of and is dedicated for use by the ANN computation system 100. The dedicated memory unit 120 allows transactions between the memory unit 120 and the controller 110 to be specifically adapted to allow a smooth and uninterrupted flow of data between the memory unit 120 and the controller 110. Such uninterrupted flow of data can significantly improve the overall throughput of the ANN computation system 100 by allowing the optoelectronic processor 140 to perform matrix multiplication at its full processing rate without being limited by slow memory operations of a conventional computer such as the computer 102. Further, because all of the data needed in performing the ANN computation can be provided by the computer 102 to the ANN computation system 100 in a single transaction, the ANN computation system 100 is able to perform its ANN computation in a self-contained manner independent of the computer 102. This self-contained operation of the ANN computation system 100 offloads the computation burden from the computer 102 and removes external dependencies in the operation of the ANN computation system 100, improving the performances of both the system 100 and the computer 102.
Example implementations of the optoelectronic processor 140 and the MC unit 130 will now be described. In some implementations, the optoelectronic processor 140 includes a laser unit 142, a modulator array 144, an optoelectronic matrix multiplication (OMM) unit 150, and an analog electronic unit 146. The modulator array 144 modulates optical waves received from the laser unit 142 to provide an optical input vector of length N based on encoded input data (e.g., digital input data) provided by the MC unit 130, and the optical input vector propagates to the OMNI unit 150. The OMM unit 150 receives the optical input vector of length N and performs, in the optical domain, a matrix multiplication on the received optical input vector. The matrix multiplication can be an N×N matrix multiplication that is determined by an internal configuration of the OMM unit 150. The internal configuration of the OMNI unit 150 can be controlled by electrical signals, such as those generated by the MC unit 130.
The OMNI unit 150 can be implemented in various ways. For example, optical modulation that is used by the modulator array 144 and/or the OMM unit 150 can include a plurality of interconnected Mach-Zehnder interferometers (MZIs).
The optical input vector is generated through the laser unit 142 and the modulator array 144. The optical input vector of length N has N independent optical signals that each have an intensity that corresponds to the value of respective element of a digital input vector of length N. As an example, the laser unit 142 can generate N light outputs. In this example, the N light outputs are of the same wavelength, and are coherent. Optical coherence of the light outputs allow the light outputs to optically interfere with each other, which is a property utilized by the OMNI unit 150 (e.g., in the operation of the MZIs). Further, the light outputs of the laser unit 142 can be substantially identical to each other. For example, the N light outputs can be substantially uniform in their intensities (e.g., within 5%, 3%, 1%, 0.5%, 0.1% or 0.01%) and in their relative phases (e.g., within 10 degrees, 5 degrees, 3 degrees, 1 degree, 0.1 degree). The uniformity of the light outputs can improve the faithfulness of the optical input vector to the digital input vector, improving the overall accuracy of the optoelectronic processor 140. In some implementations, the light outputs of the laser unit 142 can have optical powers that range from 0.1 mW to 50 mW per output, wavelengths in the near infrared range (e.g., between 900 nm and 1600 nm), and linewidths less than 1 nm. The light outputs of the laser unit 142 can be single transverse-mode light outputs.
In some implementations, the laser unit 142 includes a single laser source and an optical power splitter. The single laser source is configured to generate laser light. The optical power splitter is configured to split the light generated by the laser source into N light outputs of substantially equal intensities and phase. By splitting a single laser output into multiple outputs, optical coherence of the multiple light outputs can be achieved. The single laser source can be, for example, a semiconductor laser diode, a vertical-cavity surface-emitting laser (VCSEL), a distributed feedback (DFB) laser, or a distributed Bragg reflector (DBR) laser. The optical power splitter can be, for example, a 1:N multimode interference (MMI) splitter, a multi-stage splitter including multiple 1:2 MMI splitter or directional-couplers, or a star coupler. In some other implementations, a master-slave laser configuration can be used, in which the slave lasers are injection locked by the master laser to have a stable phase relationship to the master laser.
The light outputs of the laser unit 142 are coupled to the modulator array 144. The modulator array 144 is configured to receive the light inputs from the laser unit 142 and modulate the intensities of the received light inputs based on modulator control signals, which are electrical signals. Examples of modulators include Mach-Zehnder Interferometer (MZI) modulators, ring resonator modulators, and electro-absorption modulators. For example, an electro-absorption modulator includes electrodes that apply an electric field to a waveguide segment to modulate the absorption spectrum of the waveguide segment and thereby modulate the intensity of light propagating in the waveguide segment. The modulator array 144 has N modulators that each receives one of the N light outputs of the laser unit 142. A modulator receives a control signal that corresponds to an element of the digital input vector and modulates the intensity of the light. The control signal can be generated by the MC unit 130.
The MC unit 130 is configured to generate multiple modulator control signals and to generate multiple weight control signals under the control of the controller 110. For example, the MC unit 130 receives, from the controller 110, a first modulator control signal that corresponds to the digital input vectors to be processed by the optoelectronic processor 140. The MC unit 130 generates, based on the first modulator control signal, the modulator control signals, which are analog signals suitable for driving the modulator array 144 and the OMNI 150. The analog signals can be voltages or currents, for example, depending on the technology and design of the modulators of the array 144 and the OMNI 150. The voltages can have an amplitude that ranges from, e.g., ±0.1 V to ±10 V, and the current can have an amplitude that ranges from, e.g., 100 μA to 100 mA. In some implementations, the MC unit 130 can include modulator drivers that are configured to buffer, amplify, or condition the analog signals so that the modulators of the array 144 and the OMNI 150 can be adequately driven. For example, some types of modulators can be driven with a differential control signal. In such cases, the modulator drivers can be differential drivers that produce a differential electrical output based on a single-ended input signal.
In some implementations, the optoelectronic processor 140 can include some types of modulators that have a 3 dB bandwidth that is less than a desired processing rate of the optoelectronic processor 140. In such cases, the modulator drivers can include pre-emphasis circuits or other bandwidth-enhancing circuits that are designed to extend the operating bandwidth of the modulators. Such bandwidth-enhancement can be useful, for example, with modulators that are based on PIN diode structures forward-biased to use carrier injection for modulating a refractive index of a portion of a waveguide that is guiding an optical wave being modulated. For example, if the modulator is an MZI modulator, the PIN diode structure can be used to implement a phase shifter in one or both arms of the MZI modulator. Configuring the phase shifter for forward-biased operation facilitates shorter modulator lengths and more compact overall design. Either one or both of the modulator array 144 and the OMM unit 150 can include modulators that are based on PIN diode structures forward-biased to use carrier injection for modulation. This is useful for making the modulator array 144 and/or the OMNI unit 150 (which can have a large number of modulators) more compact.
In some implementations, either one or both of the modulator array 144 and the OMNI unit 150 can use modulators that use carrier accumulation designs that incorporate MOS-type capacitor structures into the optical waveguides, e.g., MOS-capacitor-based MZI modulators. The selection of which type of modulator to use, such as whether to use carrier injection or carrier accumulation, can be based in part on the overall semiconductor manufacturing process used to fabricate the optoelectronic processor 140.
For example, in a pre-emphasis form of bandwidth-enhancement, an analog electrical signal (e.g., voltage or current) that drives a modulator can be shaped to include a transient pulse that overshoots a change in an analog signal level that represents a given digital data value of a modulator control signal in a series of digital data values. Each digital data value can have any number of bits, including a single 1-bit data value, as assumed for the rest of this example. Thus, if a value of a bit is the same as a previous value, the analog electrical signal driving a modulator is maintained at a steady-state level (e.g., a signal level X0 for a bit value of 0, and a higher signal level X1 for a bit value of 1). However, if a bit changes from 0 to 1, the corresponding analog electrical signal used to drive the modulator can include a transient pulse with a peak value of X1+(X1−X0) at the onset of the bit transition before leveling off to a steady state value of X1. Likewise, if a bit changes from 1 to 0, the corresponding analog electrical signal used to drive the modulator can include a transient pulse with a peak value of X0+(X0−X1) at the onset of the bit transition before leveling off to a steady state value of X0. The size and length of the transient pulse can be selected to optimize the bandwidth enhancement (e.g., maximizing an open area of an eye diagram of a non-return-to-zero (NRZ) modulation pattern).
In a charge-pump form of bandwidth enhancement, an analog current signal that drives a modulator can be shaped to include a transient pulse that moves a precisely determined amount of charge.
The value of the voltage VCP can be tuned before operation such that a nominal charge Q stored in the charge pump capacitor 4402 is precisely calibrated based on a measured value of the capacitance Cp (which can have some variability due to uncertainties during manufacturing, for example). For example, the voltage VCP can be equal to the nominal charge Q divided by the capacitance Cp. The resulting change in the refractive index of a portion of a waveguide intersecting the PIN diode can then provide a shift in phase of a guided optical wave that is linearly proportional to the amount of charge Q that is moved between the PIN diode (e.g., stored via the internal capacitance Cd) and the charge pump capacitor 4402. If the driving voltage is changing from a low value to a high value, an inflow of current from the charge pump capacitor 4402 to the PIN diode delivers a predetermined quantity of charge in a short amount of time (i.e., the integral of the positive current over time). If the driving voltage is changing from a high value to a low value, an outflow of current from the PIN diode to the charge pump capacitor 4402 removes a predetermined quantity of charge in a short amount of time (i.e., the integral of the negative current over time). After this relatively short switching time, a steady state current is provided by a current source 4412, controlled by a switch 4414, to replace the charge that was lost due to the internal capacitor losing current through the internal resistance R while the driving voltage is held (e.g., during a hold time of a particular digital value). The use of such a charge-pump configuration can have advantages such as better precision over other techniques (including some pre-emphasis techniques) since the amount of charge that moves in the short switching time is dependent on a constant physical parameter (Cp) and a steady state control value (VCP), and therefore is precisely controllable and repeatable.
Other forms of bandwidth enhancement are also possible. For example, a matching electrical circuit can be used to shape the amplitude changes in electrical signals applied to the semiconductor diode modulating an optical wave. A control electrical signal can be applied to the semiconductor diode through a matching electrical circuit that is configured to match an impedance associated with the semiconductor diode without significantly changing an amplitude of the applied electrical signal (e.g., without introducing the amplitude changes of pre-emphasis/de-emphasis).
In some implementations, the matching electrical circuit is a passive electrical circuit, such as a circuit that consists essentially of an inductor. For example, one terminal of the inductor is connected to a voltage source and another terminal of the inductor is connected to a terminal of the modulator circuit 4400. An advantage of an inductor over a passive matching electrical circuit that uses, for example, an RC network of resistor(s) and capacitor(s) is consistency of fabrication. Design of a suitable RC network typically relies on cancellation between a pole and a corresponding zero in a transfer function to make the transfer function close to one, but the design tolerances needed to achieve the cancellation make the device more susceptible to fabrication errors. For example, the cancellation would require fine-tuning the products of different resistance and capacitance values to be equal to each other. Instead, by using an inductor with an appropriate value of inductance (e.g., 1 to 10 nH), the matching can be achieved without such fine-tuning in way that is more tolerant to fabrication errors.
In some cases, the modulators of the array 144 and/or the OMNI 150 can have nonlinear transfer functions. For example, an MZI optical modulator can have a nonlinear relationship (e.g., a sinusoidal dependence) between the applied control voltage and its transmission. In such cases, the first modulator control signals can be adjusted, or compensated, based on the nonlinear transfer function of the modulators such that a linear relationship between the digital input vectors and the generated optical input vectors can be maintained. Maintaining such linearity is typically important in ensuring that the input to the OMNI unit 150 is an accurate representation of the digital input vector. In some implementations, the compensation of the first modulator control signal can be performed by the controller 110 by a lookup table that maps a value of the digital input vector to a value to be output by the MC unit 130 such that the resulting modulated optical signals are linearly proportional to the elements of the digital input vector. The lookup table can be generated by characterizing the nonlinear transfer function of the modulator and calculating an inverse function of the nonlinear transfer function.
In some implementations, the nonlinearity of the modulators and resulting nonlinearity in the generated optical input vectors can be compensated by ANN computation algorithms.
The optical input vector generated by the modulator array 144 is input to the OMNI unit 150. The optical input vector can be N spatially separated optical signals that each have an optical power corresponding to the elements of the digital input vector. The optical power of the optical signals typically range from, e.g., 1 μW to 10 mW. The OMNI unit 150 receives the optical input vector and performs a matrix multiplication based on its internal configuration. The internal configuration is controlled by electrical signals generated by the MC unit 130. For example, the MC unit 130 receives, from the controller 110, a second modulator control signal that corresponds to the neural network weights to be implemented by the OMM unit 150. The MC unit 130 generates, based on the second modulator control signal, the weight control signals, which are analog signals suitable for controlling the reconfigurable elements within the OMM unit 150. The analog signals can be voltages or currents, for example, depending on the type of the reconfiguring elements of the OMNI unit 150. The voltages can have an amplitude that ranges from, e.g., 0.1 V to 10 V, and the current can have an amplitude that ranges from, e.g., 100 μA to 10 mA.
The modulator array 144 can operate at a modulation rate that is different from a reconfiguration rate at which the OMNI unit 150 can be reconfigured. The optical input vector generated by the modulator array 144 propagates through the OMM unit at a substantial fraction of the speed of light (e.g., 80%, 50%, or 25% of the speed of light), depending on the optical properties (e.g., effective index) of the OMNI unit 150. For a typical OMNI unit 150, the propagation time of the optical input vector is in the range of 1 to 10's of picoseconds, which corresponds to 10's to 100's of GHz in processing rate. As such, the rate at which the optoelectronic processor 140 can perform matrix multiplication operations is limited in part by the rate at which the optical input vector can be generated. Modulators having bandwidths of 10's of GHz are readily available, and modulators having bandwidth exceeding 100 GHz are being developed. As such, the modulation rate of the modulator array 144 can range, for example, from 5 GHz, 8 GHz, or 10's of GHz to 100's of GHz. In order to sustain the operation of the modulator array 144 at such modulation rate, the integrated circuitry of the controller 110 can be configured to output control signals for the MC unit 130 at a rate greater than or equal to, for example, 5 GHz, 8 GHz, 10 GHz, 20 GHz, 25 GHz, 50 GHz, or 100 GHz.
The reconfiguration rate of the OMNI unit 150 can be significantly slower than the modulation rate depending on the type of the reconfigurable elements implemented by the OMNI unit 150. For example, the reconfigurable elements of the OMNI unit 150 can be a thermo-optic type that uses a micro-heater to adjust a temperature of an optical waveguide of the OMM unit 150, which in turn affects the phase of an optical signal within the OMM unit 150 and leads to matrix multiplication. Due to the thermal time constants associated with heating and cooling of structures, the reconfiguration rate can be limited to 100's of kHz to 10's of MHz, for example. As such, the modulator control signals for controlling the modulator array 144 and the weight control signals for reconfiguring the OMM unit 150 can have significantly different requirements in speed. Further, the electrical characteristics of the modulator array 144 can differ significantly from those of the reconfigurable elements of the OMM unit 150.
To accommodate the different characteristics of the modulator control signals and the weight control signals, in some implementations, the MC unit 130 can include a first MC subunit 132, and a second MC subunit 134. The first MC subunit 132 can be specifically configured to generate the modulator control signals for the input vector, and the second MC subunit 134 can be specifically configured to generate weight control signals for the matrix multiplication. For example, the modulation rate of the modulator array 144 can be 25 GHz, and the first MC subunit 132 can have a per-channel output update rate of 25 giga-samples per second (GSPS) and a resolution of 8 bits or higher. The reconfiguration rate of the OMNI unit 150 can be 1 MHz, and the second MC subunit 134 can have an output update rate of 1 mega-samples per second (MSPS) and a resolution of 10 bits. Implementing separate MC subunits 132 and 134 allows independent optimization of the MC subunits for respective signals, which can reduce the total power consumption, complexity, cost, or combination thereof of the MC unit 130. It should be noted that while the MC subunits 132 and 134 are described as sub elements of the MC unit 130, in general, the MC subunits 132 and 134 can be integrated on a common chip, or be implemented as separate chips.
Based on the different characteristics of the first MC subunit 132 and the second MC subunit 134, in some implementations, the memory unit 120 can include a first memory subunit and a second memory subunit. The first memory subunit can be a memory dedicated to storing of the input dataset and the digital input vectors, and can have an operating speed sufficient to support the modulation rate. The second memory subunit can be a memory dedicated to storing of the neural network weights, and can have an operation speed sufficient to support the reconfiguration rate of the OMNI unit 150. In some implementations, the first memory subunit can be implemented using SRAM and the second memory subunit can be implemented using DRAM. In some implementations, the first and second memory subunits can be implemented using DRAM. In some implementations, the first memory subunit can be implemented as a part of or as a cache of the controller 110. In some implementations, the first and second memory subunits can be implemented by a single physical memory device as different address spaces.
The OMNI unit 150 outputs an output vector of length N, which corresponds to the result of the N×N matrix multiplication of the optical input vector and the neural network weights. In some implementations, the output vector can comprise electrical signals (e.g., voltages or currents), and in other implementations the output vector can comprise optical signals. The OMM unit 150 is coupled to the analog electronic unit 146, which is configured to perform any analog electronic processing for implementations in which the output vector is an optical output vector, can also be configured to perform optical-to-electrical conversion. For example, the analog electronic unit 146 can include an array of N photodetectors configured to absorb the optical signals and generate photocurrents, and an array of N transimpedance amplifiers configured to convert the photocurrents into the output voltages. Alternatively, if an electrical output vector is received from the OMM unit 150, there can be photodetectors and transimpedance amplifiers within the OMNI unit 150. The bandwidths of the photodetectors and the transimpedance amplifiers can be set based on the modulation rate of the modulator array 144. The photodetectors can be formed from various materials based on the wavelengths of the optical output vector being detected. Examples of the materials for photodetectors include germanium, silicon-germanium alloy, and indium gallium arsenide (InGaAs).
The analog electronic unit 146 is coupled to the ADC unit 160. The ADC unit 160 is configured to convert the N electrical signals output from the analog electronic unit 146 into N digitized optical outputs, which are quantized digital representations of the output voltages. For example, the ADC unit 160 can be an N channel ADC. The controller 110 can obtain, from the ADC unit 160, the N digitized optical outputs corresponding to the optical output vector of the optical matrix multiplication unit 150. The controller 110 can form, from the N digitized optical outputs, a digital output vector of length N that corresponds to the result of the N×N matrix multiplication of the input digital vector of length N. In some implementations, if no analog electronic processing is needed and the OMM unit 150 provides electrical output signals, the analog electronic unit 146 can be omitted and the OMNI unit 150 can be connected directly to the ADC unit 160.
Various electrical components of the ANN computation system 100 can be integrated in various ways. For example, the controller 110 can be an application specific integrated circuit that is fabricated on a semiconductor die. Other electrical components, such as the memory unit 120, the MC unit 130, the ADC unit 160, or combination thereof can be monolithically integrated on the semiconductor die on which the controller 110 is fabricated. As another example, two or more electrical components can be integrated as a System-on-Chip (SoC). In a SoC implementation, the controller 110, the memory unit 120, the MC unit 130, and the ADC unit 160 can be fabricated on respective dies, and the respective dies can be integrated on a common platform (e.g., an interposer) that provides electrical connections between the integrated components. Such SoC approach can allow faster data transfer between the electronic components of the ANN computation system 100 relative to an approach where the components are separately placed and routed on a printed circuit board (PCB), thereby improving the operating speed of the ANN computation system 100. Further, the SoC approach can allow use of different fabrication technologies optimized for different electrical components, which can improve the performance of the different components and reduce overall costs over a monolithic integration approach. While the integration of the controller 110, the memory unit 120, the MC unit 130, and the ADC unit 160 has been described, in general, a subset of the components can be integrated while other components are implemented as discrete components for various reasons, such as performance or cost. For example, in some implementations, the memory unit 120 can be integrated with the controller 110 as a functional block within the controller 110.
Various optical components of the ANN computation system 100 can also be integrated in various ways. Examples of the optical components of the ANN computation system 100 include the laser unit 142, the modulator array 144, the OMM unit 150, and the photodetectors of the analog electronic unit 146. These optical components can be integrated in various ways to improve performance and/or reduce cost. For example, the laser unit 142, the modulator array 144, the OMM unit 150, and the photodetectors can be monolithically integrated on a common semiconductor substrate as a photonic integrated circuit (PIC). On a photonic integrated circuit formed based on a compound semiconductor material system (e.g., III-V compound semiconductors such as InP), lasers, modulators such as electro-absorption modulators, waveguides, and photodetectors can be monolithically integrated on a single die. Such monolithic integration approach can reduce the complexities of aligning the inputs and outputs of various discrete optical components, which can require alignment accuracies ranging from sub-micron to a few microns. As another example, the laser source of the laser unit 142 can be fabricated on a compound-semiconductor die, while the optical power splitter of the laser unit 142, the modulator array 144, the OMNI unit 150, and the photodetectors of the analog electronic unit 146 can be fabricated on a silicon die. PICs fabricated on a silicon wafer, which can be referred to as silicon photonics technology, typically has a greater integration density, higher lithographic resolution, and lower cost relative to the III-V based PICs. Such greater integration density can be beneficial in fabrication of the OMM unit 150, as the OMM unit 150 typically includes 10's to 100's of optical components such as power splitters and phase shifters. Further, the higher lithographic resolution of the silicon photonics technology can reduce fabrication variation of the OMM unit 150, improving the accuracy of the OMNI unit 150.
The ANN computation system 100 can be implemented in a variety of form factors. For example, the ANN computation system 100 can be implemented as a co-processor that is plugged into a host computer. Such system 100 can have, for example, a form factor of a PCI express card and communicate with the host computer over the PCIe bus. The host computer can host multiple co-processor type ANN computation systems 100, and be connected to the computer 102 over a network. This type of implementation can be suitable for use in a cloud datacenter where racks of servers can be dedicated to processing ANN computation requests received from other computers or servers. As another example, the co-processor type ANN computation system 100 can be plugged directly into the computer 102 issuing the ANN computation requests.
In some implementations, the controller 110, the memory unit 120, the modulator control unit 130, the ADC unit 160, and a microprocessor can be monolithically integrated on a semiconductor die. In some implementations, the controller 110, the memory unit 120, the modulator control unit 130, the ADC unit 160, the microprocessor, and the system main memory can be integrated as system-on-chip. For example, this allows the artificial neural network computation system 100 to be used in a portable device, such as a laptop computer, a tablet computer, or a mobile phone. The microprocessor can include, e.g., multiple high-performance processor cores, multiple high-efficiency processor cores, multiple graphics processors, multiple electronic neural engine cores, level 1 cache, and level 2 cache. The microprocessor can use the electronic neural engine cores to execute artificial neural network computation instructions that have been optimized for conventional electronic neural engine cores, and use the ANN computation system 100 to execute artificial neural network computation instructions that have been optimized for the optical processing performed by the optoelectronic processor 140. The microprocessor can be, e.g., a reduced instruction set computer or a complex instruction set computer. The operating system can be designed to take into account the ANN computation system 100, e.g., powering up the ANN computation system 100 to perform specific tasks that are more suitable to be performed by the ANN computation system 100, and powering down the ANN computation system 100 or placing the ANN computation system 100 in a standby mode when not performing such tasks, thereby achieving overall high computation performance and low power consumption.
In some implementations, the ANN computation system 100 can be integrated onto a physical system that requires real-time ANN computation capability. For example, systems that rely heavily on real-time artificial intelligence tasks such as autonomous vehicles, autonomous drones, object- or face-recognizing security cameras, and various Internet-of-Things (IoT) devices can benefit from having ANN computation system 100 directly integrated with other subsystems of such systems. Having directly-integrated ANN computation system 100 can enable real-time artificial intelligence in devices with poor or no internet connectivity, and enhance the reliability and availability of mission-critical artificial intelligence systems.
While the MC unit 130 and the ADC unit 160 are illustrated to be coupled to the controller 110, in some implementations, the MC unit 130, the ADC unit 160 or both can alternatively, or additionally, be coupled to the memory unit 120. For example, a direct memory access (DMA) operation by the MC unit 130 or the ADC unit 160 can reduce the computation burden on the controller 110 and reduce latency in reading from and writing to the memory unit 120, further improving the operating speed of the ANN computation system 100.
At 210, an artificial neural network (ANN) computation request comprising an input dataset and a first plurality of neural network weights is received. The input dataset includes a first digital input vector. The first digital input vector is a subset of the input dataset. For example, it can be a sub-region of an image. The ANN computation request can be generated by various entities, such as the computer 102. The computer can include one or more of various types of computing devices, such as a personal computer, a server computer, a vehicle computer, and a flight computer. The ANN computation request generally refers to an electrical signal that notifies or informs the ANN computation system 100 of an ANN computation to be performed. In some implementations, the ANN computation request can be divided into two or more signals. For example, a first signal can query the ANN computation system 100 to check whether the system 100 is ready to receive the input dataset and the first plurality of neural network weights. In response to a positive acknowledgement by the system 100, the computer can send a second signal that includes the input dataset and the first plurality of neural network weights.
At 220, the input dataset and the first plurality of neural network weights are stored. The controller 110 can store the input dataset and the first plurality of neural network weights in the memory unit 120. Storing of the input dataset and the first plurality of neural network weights in the memory unit 120 can allow flexibilities in the operation of the ANN computation system 100 that, for example, can improve the overall performance of the system. For example, the input dataset can be divided into digital input vectors of a set size and format by retrieving desired portions of the input dataset from the memory unit 120. Different portions of the input dataset can be processed in various order, or be shuffled, to allow various types of ANN computations to be performed. For example, shuffling can allow matrix multiplication by block matrix multiplication technique in cases where the input and output matrix sizes are different. As another example, storing of the input dataset and the first plurality of neural network weights in the memory unit 120 can allow queuing of multiple ANN computation requests by the ANN computation system 100, which can allow the system 100 to sustain operation at its full speed without periods of inactivity.
In some implementations, the input dataset can be stored in the first memory subunit, and the first plurality of neural network weights can be stored in the second memory subunit.
At 230, a first plurality of modulator control signals are generated based on the first digital input vector, and a first plurality of weight control signals are generated based on the first plurality of neural network weights. The controller 110 can send a first modulator control signal to the MC unit 130 for generating the first plurality of modulator control signals. The MC unit 130 generates the first plurality of modulator control signals based on the first modulator control signal, and the modulator array 144 generates the optical input vector representing the first digital input vector.
The first modulator control signal can include multiple digital values to be converted by the MC unit 130 into the first plurality of modulator control signals. The multiple digital values are generally in correspondence with the first digital input vector, and can be related through various mathematical relationships or look-up tables. For example, the multiple digital values can be linearly proportional to the values of the elements of the first digital input vector. As another example, the multiple digital values can be related to the elements of the first digital input vector through a look-up table configured to maintain a linear relationship between the digital input vector and the optical input vector generated by the modulator array 144.
The controller 110 can send a second modulator control signal to the MC unit 130 for generating the first plurality of weight control signals. The MC unit 130 generates the first plurality of weight control signals based on the second modulator control signal, and the OMNI unit 150 is reconfigured according to the first plurality of weight control signals, implementing a matrix corresponding to the first plurality of neural network weights.
The second modulator control signal can include multiple digital values to be converted by the MC unit 130 into the first plurality of weight control signals. The multiple digital values are generally in correspondence with the first plurality of neural network weights, and can be related through various mathematical relationships or look-up tables. For example, the multiple digital values can be linearly proportional to the first plurality of neural network weights. As another example, the multiple digital values can be calculated by performing various mathematical operations on the first plurality of neural network weights to generate weight control signals that can configure the OMNI unit 150 to perform a matrix multiplication corresponding to the first plurality of neural network weights.
At 240, a first plurality of digitized optical outputs corresponding to the optical output vector of the optical matrix multiplication unit is obtained. The optical input vector generated by the modulator array 144 is processed by the OMNI unit 150 and transformed into an optical or electrical output vector. If the output vector is an optical output vector, the optical output vector is detected by the analog electronic unit 146 and converted into electrical signals that can be converted into digitized values by the ADC unit 160. The controller 110 can, for example, send a conversion request to the ADC unit 160 to begin a conversion of the voltages output by the analog electronic unit 146 into digitized optical outputs. Once the conversion is complete, the ADC unit 160 can send the conversion result to the controller 110. Alternatively, the controller 110 can retrieve the conversion result from the ADC unit 160. The controller 110 can form, from the digitized optical outputs, a digital output vector that corresponds to the result of the matrix multiplication of the input digital vector. For example, the digitized optical outputs can be organized, or concatenated, to have a vector format.
In some implementations, the ADC unit 160 can be set or controlled to perform an ADC conversion based on a modulator control signal issued to the MC unit 130 by the controller 110. For example, the ADC conversion can be set to begin at a preset time following the generation of the modulation control signal by the MC unit 130. Such control of the ADC conversion can simplify the operation of the controller 110 and reduce the number of necessary control operations.
At 250, a nonlinear transformation is performed on the first digital output vector to generate a first transformed digital output vector. A node, or an artificial neuron, of an ANN operates by first performing a weighted sum of the signals received from nodes of a previous layer, then performing a nonlinear transformation (“activation”) of the weighted sum to generate an output. Various types of ANN can implement various types of differentiable, nonlinear transformations. Examples of nonlinear transformation functions include a rectified linear unit (RELU) function, a Sigmoid function, a hyperbolic tangent function, an X2 function, and a |X| function. Such nonlinear transformations are performed on the first digital output by the controller 110 to generate the first transformed digital output vector. In some implementations, the nonlinear transformations can be performed by a specialized digital integrated circuitry within the controller 110. For example, the controller 110 can include one or more modules or circuit blocks that are specifically adapted to accelerate the computation of one or more types of nonlinear transformations.
At 260, the first transformed digital output vector is stored. The controller 110 can store the first transformed digital output vector in the memory unit 120. In cases where the input dataset is divided into multiple digital input vectors, the first transformed digital output vector corresponds to a result of the ANN computation of a portion of the input dataset, such as the first digital input vector. As such, storing of the first transformed digital output vector allows the ANN computation system 100 to perform and store additional computations on other digital input vectors of the input dataset to later be aggregated into a single ANN output.
At 270, an artificial neural network output generated based on the first transformed digital output vector is output. The controller 110 generates an ANN output, which is a result of processing the input dataset through the ANN defined by the first plurality of neural network weights. In cases where the input dataset is divided into multiple digital input vectors, the generated ANN output is an aggregated output that includes the first transformed digital output, but can further include additional transformed digital outputs that correspond to other portions of the input dataset. Once the ANN output is generated, the generated output is sent to a computer, such as the computer 102, that originated the ANN computation request.
Various performance metrics can be defined for the ANN computation system 100 implementing the process 200. Defining performance metrics can allow a comparison of performance of the ANN computation system 100 that implements the optoelectronic processor 140 with other systems for ANN computation that instead implement electronic matrix multiplication units. In one aspect, the rate at which an ANN computation can be performed can be indicated in part by a first loop period defined as a time elapsed between the step 220 of storing, in the memory unit, the input dataset and the first plurality of neural network weights, and the step 260 of storing, in the memory unit, the first transformed digital output vector. This first loop period therefore includes the time taken in converting the electrical signals into optical signals (e.g., step 230), performing the matrix multiplication in the optical domain, and converting the result back into the electrical domain (e.g., step 240). Steps 220 and 260 both involves storing of data into the memory unit 120, which are steps shared between the ANN computation system 100 and conventional ANN computation systems without the optoelectronic processor 140. As such, the first loop period measuring the memory-to-memory transaction time can allow a realistic or fair comparison of ANN computation throughput to be made between the ANN computation system 100 and ANN computation systems without the optoelectronic processor 140, such as systems implementing electronic matrix multiplication units.
Due to the rate at which the optical input vectors can be generated by the modulator array 144 (e.g., at 25 GHz) and the processing rate of the OMNI unit 150 (e.g., greater than 100 GHz), the first loop period of the ANN computation system 100 for performing a single ANN computation of a single digital input vector can approach the reciprocal of the speed of the modulator array 144, e.g., 40 ps. After accounting for latencies associated with the signal generation by the MC unit 130 and the ADC conversion by the ADC unit 160, the first loop period can, for example, be less than or equal to 100 ps, less than or equal to 200 ps, less than or equal to 500 ps, less than or equal to 1 ns, less than or equal to 2 ns, less than or equal to 5 ns, or less than or equal to 10 ns.
As a comparison, execution time of a multiplication of an M×1 vector and an M×M matrix by an electronic matrix multiplication unit is typically proportional to M2−1 processor clock cycles. For M=32, such multiplication would take approximately 1024 cycles, which at 3 GHz clock speed results in an execution time exceeding 300 ns, which is orders of magnitude slower than the first loop period of the ANN computation system 100.
In some implementations, the process 200 further includes a step of generating a second plurality of modulator control signals based on the first transformed digital output vector. In some types of ANN computations, a single digital input vector can be repeatedly propagated through, or processed by, the same ANN. An ANN that implements multi-pass processing can be referred to as a recurrent neural network (RNN). A RNN is a neural network in which the output of the network during a (k)th pass through the neural network is recirculated back to the input of the neural network and used as the input during the (k+1)th pass. RNNs can have various applications in pattern recognition tasks, such as speech or handwriting recognition. Once the second plurality of modulator control signals are generated, the process 200 can proceed from step 240 through step 260 to complete a second pass of the first digital input vector through the ANN. In general, the recirculation of the transformed digital output to the digital input vector can be repeated for a preset number of cycles depending on the characteristics of the RNN received in the ANN computation request.
In some implementations, the process 200 further includes a step of generating a second plurality of weight control signals based on a second plurality of neural network weights. In some cases, the artificial neural network computation request further includes a second plurality of neural network weights. In general, an ANN has one or more hidden layers in addition to the input and output layers. For ANN with two hidden layers, the second plurality of neural network weights can correspond, for example, to the connectivity between the first layer of the ANN and the second layer of the ANN. To process the first digital input vector through the two hidden layers of the ANN, the first digital input vector can first be processed according to the process 200 up to step 260, at which the result of processing the first digital input vector through the first hidden layer of the ANN is stored in the memory unit 120. The controller 110 then reconfigures the OMNI unit 150 to perform the matrix multiplication corresponding to the second plurality of neural network weights associated with the second hidden layer of the ANN. Once the OMM unit 150 is reconfigured, the process 200 can generate the plurality of modulator control signals based on the first transformed digital output vector, which generates an updated optical input vector corresponding to the output of the first hidden layer. The updated optical input vector is then processed by the reconfigured OMM unit 150 which corresponds to the second hidden layer of the ANN. In general, the described steps can be repeated until the digital input vector has been processed through all hidden layers of the ANN.
As previously described, in some implementations of the OMNI unit 150, the reconfiguration rate of the OMM unit 150 can be significantly slower than the modulation rate of the modulator array 144. In such cases, the throughput of the ANN computation system 100 can be adversely impacted by the amount of time spent in reconfiguring the OMNI unit 150 during which ANN computations cannot be performed. To mitigate the impact of the relatively slow reconfiguration time of the OMNI unit 150, batch processing techniques can be utilized in which two or more digital input vectors are propagated through the OMM unit 150 without a configuration change to amortize the reconfiguration time over a larger number of digital input vectors.
To implement batch processing, in some implementations, the process 200 further includes steps of generating, through the MC unit 130, a second plurality of modulator control signals based on the second digital input vector; obtaining, from the ADC unit, a second plurality of digitized optical outputs corresponding to the optical output vector of the optical matrix multiplication unit, the second plurality of digitized optical outputs forming a second digital output vector; performing a nonlinear transformation on the second digital output vector to generate a second transformed digital output vector; and storing, in the memory unit, the second transformed digital output vector. The generating of the second plurality of modulator control signals can follow the step 260, for example. Further, the ANN output of step 270 in this case is now based on both the first transformed digital output vector and the second transformed digital output vector. The obtaining, performing, and storing steps are analogous to the steps 240 through 260.
The batch processing technique is one of several techniques for improving the throughput of the ANN computation system 100. Another technique for improving the throughput of the ANN computation system 100 is through parallel processing of multiple digital input vectors by utilizing wavelength division multiplexing (WDM). WDM is a technique of simultaneously propagating multiple optical signals of different wavelengths through a common propagation channel, such as a waveguide of the OMM unit 150. Unlike electrical signals, optical signals of different wavelengths can propagate through a common channel without affecting other optical signals of different wavelengths on the same channel. Further, optical signals can be added (multiplexed) or dropped (demultiplexed) from a common propagation channel using well-known structures such as optical multiplexers and demultiplexers.
In the context of the ANN computation system 100, multiple optical input vectors of different wavelengths can be independently generated, simultaneously propagated through the OMM unit 150, and independently detected to enhance the throughput of the ANN computation system 100. Referring to
A phase shifter is a type of optical modulator. In this document the term “modulator” depending on context can refer to, e.g., the overall modulator 3950, the phase shifter 3956, or an optical amplitude modulator.
Referring to
The first and second electrodes 3962a, 3962b can be driven by a driving circuit 3966a that is similar to the circuit shown in
The phase shifter 3956b includes a heavily doped p+ region 3968a and a heavily doped n+ region 3968b having a length 2L and formed in the substrate on two sides of a second segment of the first ridge or rib waveguide 3964a. The p+ region, the intrinsic region between the p+ and n+ regions, and the n+ region form a p+−i−n+(PIN) diode. A third electrode (anode) 3970a is electrically coupled to the p+ region 3968a, and a fourth electrode (cathode) 3970b is electrically coupled to the n+ region 3968b. The third and fourth electrodes 3970a and 3970b can be driven by a driving circuit 3966b, which is similar to the driving circuit 3966a. A control voltage waveform 4404b is provided to an inverter circuit to generate a driving voltage waveform whose amplitude can be precisely calibrated to move a predetermined amount of charge to or from the phase shifter 3956b via a pump capacitor. The driving circuit 3966b is configured move a predetermined amount of charge to or from the phase shifter 3956b to control the corresponding optical phase shift applied by the phase shifter 3956b.
The phase shifter 3956c includes a heavily doped p+ region 3972a and a heavily doped n+ region 3972b having a length L and formed in the substrate on two sides of a third segment of the first ridge or rib waveguide 3964a. The p+ region, the intrinsic region between the p+ and n+ regions, and the n+ region form a p+−i−n+(PIN) diode. A fifth electrode (anode) 3974a is electrically coupled to the p+ region 3972a, and a sixth electrode (cathode) 3974b is electrically coupled to the n+ region 3972b. The fifth and sixth electrodes 3974a and 3974b can be driven by a driving circuit 3966c, which is similar to the driving circuit 3966a. A control voltage waveform 4404c is provided to an inverter circuit to generate a driving voltage waveform whose amplitude can be precisely calibrated to move a predetermined amount of charge to or from the phase shifter 3956c via a pump capacitor. The driving circuit 3966c is configured move a predetermined amount of charge to or from the phase shifter 3956c to control the corresponding optical phase shift applied by the phase shifter 3956c.
The doped p+ and n+ regions of the phase shifters 3956a, 3956b, and 3956c have lengths 4L, 2L, and L, respectively. The driving circuits 4404a, 4404b, and 4404c provide binary signals that represent 3-bit digital input values to the phase shifters 3956a, 3956b, and 3956c, which impart optical phase shifts weighted according to a 4:2:1 ratio to the light propagating in the first, second, and third waveguide segments. For example, when a binary value of ‘1’ is applied to applied to the phase shifter 3956a and the phase shifter 3956c, the phase shifter 3956a will impart an amount of optical phase shift that is four times the amount of optical phase shift imparted by the phase shifter 3956c. Similarly, when a binary value of ‘1’ is applied to applied to the phase shifter 3956b and the phase shifter 3956c, the phase shifter 3956b will impart an amount of optical phase shift that is twice the amount of optical phase shift imparted by the phase shifter 3956c. This way, the MZI 3950 converts the 3-bit digital input value to an analog output signal having 3-bit precision without the use of a digital-to-analog converter circuitry, which would be needed if the MZI 3900 of
As described above, in some examples of ANN computations, such as when batch processing techniques are used, the modulator array 144 is reconfigured at a faster rate than the rate at which the OMM unit 150 is reconfigured. For example, the modulator array 144 can use a plurality of the Mach-Zehnder modulators 3950 having the segmented design for modulating the optical waves received from the laser unit 142 to provide an optical input vector of length N based on encoded input data (e.g., digital input data) provided by the first MC subunit 132, and the optical input vector propagates to the OMM unit 150. Because the OMNI unit 150 can be reconfigured at a slower rate, in some examples the OMM unit 150 include a plurality of Mach-Zehnder modulators 3900 that do not use the segmented design. The circuitry for driving the MZI 3900 can be simpler than the circuitry for driving the multiple segments of the MZI 3950. The OMM unit 150 can have a large number of MZIs, so using the MZIs 3900 in the OMM unit 150 can reduce the complexity of the driving circuitry in the OMM unit 150.
In some implementations, the optoelectronic processor 140 is configured to perform ANN computations that require the OMM unit 150 to also be updated at a fast rate comparable to that of the modulator array 144. In this case, the OMNI unit 150 can also use a plurality of MZIs 3950 that use the segmented design.
In some implementations, the modulator array 144 includes a plurality of MZIs 3900 that do not use the segmented design, and the OMNI unit 150 includes a plurality of MZIs 3950 that use the segmented design.
In some implementations, the phase shifters within an optical modulator can have the same length, and the charge-pump bandwidth enhancement circuit pumps different charges to provide different levels of modulation? The amount of charges pumped to the diode section depends in part on the charge pump control voltage VCp and the capacitance of the pump capacitor Cp 4402. By using different charge pump control voltages VCp and/or pump capacitors Cp 4402 with different capacitances for the different diode sections in the same optical modulator, the phase shifters can apply binary weighted phase shift contributions to the optical signals being modulated.
For example, in some implementations, the phase shifters 3956a, 3956b, 3956c can all have the same length, but the charge pump circuit 3966a pumps more charges to the phase shifter 3956a than the charges pumped by the charge pump circuit 3966b to the phase shifter 3956b, and the charge pump circuit 3966b pumps more charges to the phase shifter 3956b than the charges pumped by the charge pump circuit 3966c to the phase shifter 3956c, comparable results can be achieved as using phase shifters of different lengths. For example, the charge pump circuit 3966a uses charge-pump control voltage VCp1, the charge pump circuit 3966b uses charge-pump control voltage VCp2, the charge pump circuit 3966c uses charge-pump control voltage VCp3, and VCp1>VCp2>VCp3. For example, the charge pump circuit 3966a uses a charge pump capacitor having a capacitance Cp1, the charge pump circuit 3966b uses a charge pump capacitor having a capacitance Cp2, the charge pump circuit 3966c uses a charge pump capacitor having a capacitance Cp3, and Cp1>Cp2>Cp3. The advantage of this design is to reduce the overall length of the modulator 3950.
In some examples, the phase shifter 3956b can be longer than the phase shifter 3956c, but less than twice the length of the phase shifter 3956c. The charge pump control voltage VCp2 and/or the charge pump capacitance Cp2 of the phase shifter 3956b is larger than the charge pump control voltage VCp3 and/or the charge pump capacitance Cp3 of the phase shifter 3956c, such that in response to a digital input value of ‘1’, the phase shift imparted by the phase shifter 3956b to the optical wave traveling in the waveguide 3964a is twice the phase shift imparted by the phase shifter 3956c. Similarly, the phase shifter 3956a can be longer than the phase shifter 3956b, but less than twice the length of the phase shifter 3956b. The charge pump control voltage VCp1 and/or the charge pump capacitance Cp1 of the phase shifter 3956a can be larger than the charge pump control voltage VCp2 and/or the charge pump capacitance Cp2 of the phase shifter 3956b, such that in response to a digital input value of ‘1’, the phase shift imparted by the phase shifter 3956a to the optical wave traveling in the waveguide 3964a is twice the phase shift imparted by the phase shifter 3956b. The highest charge pump control voltage VCp can depend on the semiconductor process used to fabricate the optoelectronic processor.
Referring to
Referring to
In some implementations, each digital input value has N bits, and the arm 3954a is configured to have a set of N phase shifters, including a first phase shifter having a length of 2N-1L, a second phase shifter having a length of 2N-2L, . . . , and an N-th phase shifter having a length of L.
In the examples shown in
In some implementations, the MZI having segmented design can have phase shifters on both arms (e.g., 3954a and 3954b).
In some implementations, the number of phase shifters in the arm 3954a is different from the number of bits in each digital input value. In this case, the diode sections apply different respective modulation contributions to the optical wave propagating through the optical waveguide portion to provide partial digital-to-analog conversion. Additional circuitry can be used to complete the digital-to-analog conversion.
Referring to
In general, a module that needs to convert an N-bit digital electric signal to an analog optical signal that has N-bit precisions can use p-bit DACs to drive an MZI having a set of q phase shifters in one arm, in which p×q=N. The selection of p and q depends on how fast the MZI needs to be reconfigured, and how much area is available to accommodate the multiple phase shifters of the MZI. A smaller p results in a faster reconfiguration rate, and a smaller q results in a shorter MZI.
In some implementations, the semiconductor diodes are operated in a forward-biased state to take advantage of index modulation effects of carrier injection, which can be implemented using a relatively short optical path length (e.g., less than about 1.0 mm, or less than about 0.5 mm, or less than about 0.1 mm). This short optical path length enables compact modulator arrangement within an integrated optical device that can use many MZIs or other types of modulators.
In addition to interferometric optical modulators such as the MZI 3950, non-interferometric optical modulators can be used to implement segmented design for digital-to-analog conversion. For example, absorption modulators can be used along a single optical waveguide instead of phase shifters in one or more arms of an MZI. A segmented design also facilitates use of signal conditioning for bandwidth enhancement, such as pre-emphasis and/or de-emphasis, which can be implemented independently for each segment.
The term “pre-emphasis” refers to the part in which the charge-pump circuit 4416 quickly pumps charges to the modulator circuit 4400 via the capacitor Cp 4402 to cause the waveform at the top plate of the capacitor Cp 4402 to rise quickly, e.g., 1706, to a level greater than the steady state voltage. The term “de-emphasis” refers to the part in which the charge-pump circuit 4416 quickly removes charges from the modulator circuit 4400 via the capacitor Cp 4402 to cause the waveform at the top plate of the capacitor Cp 4402 to drop quickly, e.g., 1708, to a level lower than the steady state voltage.
The charge-pump bandwidth enhancement circuit 4416 can have a number of advantages compared to the conventional pre-emphasis circuit that uses voltage driving. For example, some pre-emphasis conventional circuit can only perform pre-emphasis, whereas the charge-pump bandwidth enhancement circuit 4416 can perform both pre-emphasis and de-emphasis. For example, some conventional circuit pre-emphasis circuit that uses voltage driving requires a much higher voltage source from an external source and bias the modulator at half of the supply voltage, or requires a complex voltage boost circuit. By comparison, the charge-pump bandwidth enhancement circuit 4416 can use a lower voltage source and does not require a complex voltage boost circuit. In addition, because the conventional pre-emphasis circuit is voltage driven, in order to ensure that the modulator is not driven over a targeted phase, the convention circuit can require implementing an additional control phase to stop the emphasis circuit. By comparison, the charge-pump bandwidth enhancement circuit 4416 can move precise amounts of electric charges to or from the capacitor Cp 4402 and can accurately control the voltage applied to the modulator, and thus does not need to implement the additional control phase to stop the emphasis circuit.
In some implementations, a segmented design for digital-to-analog conversion uses an alternative arrangement of the segments that changes the position of a segment corresponding to the least significant bit to be closer to the center of the segments. Such an arrangement can mitigate the effects of changes in certain device characteristics to improve the accuracy of the digital-to-analog conversion.
The span over the device position in the illustrated variation could correspond, for example, to the span of a waveguide of the optical modulator, in which edges of the waveguide may experience more variability due to influences of the environment near the edges of the waveguide. The span over the device position can also correspond, for example, to the span of a wafer, where edges of the wafer may experience more fabrication variability, or to the span of a single die of a wafer, where there the edges of a mask pattern may experience more fabrication variability. An example of a characteristic that can affect doping concentration is “mask opening density” which changes based on the changes in the size of openings in a mask. Similarly, if the variation is due to changing temperature, there may be more variability in temperature on the edges of a device.
By using the alternative segment arrangement described herein, a segment for the LSB and other low order bits can be positioned in a location where the variance in the doping concentration is relatively small (e.g., near the center), as described in more detail below. Since the low order bits determine the achievable resolution, these techniques can potentially increase resolution of the DAC.
Using the examples described above, or other non-monotonic arrangements, the device is not affected by the variation in doping concentration as much because the center part (or near the center part) of the waveguide, with less doping variation, is reserved for the most sensitive bits (i.e., the shortest segment(s)). The bits that are less sensitive to doping variation are on the outside where the doping variation is higher. A reason the longer segments for the higher order bits are less sensitive to doping variation is because their length gives them the ability to average out more of the variation.
For example, the modulator that uses the segmented design in
In this example, the laser unit 142 outputs four light signals on four optical waveguides 4050a, 4050b, 4050c, and 4050d. The light signal on the waveguide 4050a is modulated by the optical modulator 3960 according to the first 4-bit value of the digital input vector. The light signal on the waveguide 4050b is modulated by the optical modulator 3970 according to the second 4-bit value of the digital input vector. The light signal on the waveguide 4050c is modulated by the optical modulator 3980 according to the third 4-bit value of the digital input vector. The light signal on the waveguide 4050d is modulated by the optical modulator 3990 according to the fourth 4-bit value of the digital input vector.
The optical modulator 3960 includes two 1×2 port multi-mode interference couplers (MMI 1×2) 3962a and 3962b, two balanced arms 3964a and 3964b, and four sub-modulators 3966a, 3966b, 3966c, and 3966d in one arm (or four sub-modulators in each arm). For example, each sub-modulator can include a phase shifter. The sub-modulators 3966a, 3966b, 3966c, and 3966d receive the LSB (B11), 2nd-bit (B12), 3rd-bit (B13), and MSB (B14), respectively, of the first value in the input vector.
The sub-modulator 3966a includes a first waveguide segment, the sub-modulator 3966b includes a second waveguide segment, the sub-modulator 3966c includes a third waveguide segment, and the sub-modulator 3966d includes a fourth waveguide segment. The second waveguide segment is approximately twice as long as the first waveguide segment, the third waveguide segment is approximately four times as long as the first waveguide segment, and the fourth waveguide segment is approximately eight times as long as the first waveguide segment.
The optical modulator 3970 includes two 1×2 port multi-mode interference couplers (MMI 1×2) 3972a and 3972b, two balanced arms 3974a and 3974b, and four sub-modulators 3976a, 3976b, 3976c, and 3976d in one arm (or four sub-modulators in each arm). For example, each sub-modulator can include a phase shifter. The sub-modulators 3976a, 3976b, 3976c, and 3976d receive the LSB (B21), 2nd-bit (B22), 3rd-bit (B23), and MSB (B24), respectively, of the second value in the input vector. The optical modulators 3980 and 3990 are configured in a manner similar to the optical modulators 3960 and 3970.
The optical modulator 3960 modulates the optical signal in the input waveguide 4050a according to the four bits of the first value in the digital input vector, and generates an analog optical signal 3968 that represents the first value of the digital input vector. The optical modulator 3960 converts the first 4-bit value of the digital input vector to a first analog optical signal. The optical modulator 3970 modulates the optical signal in the input waveguide 3050b according to the four bits of the second value in the digital input vector, and generates an analog optical signal 3978 that represents the second value of the digital input vector. The optical modulator 3970 converts the second 4-bit value of the digital input vector to a second analog optical signal.
The optical modulator 3980 modulates the optical signal in the input waveguide 3050c according to the four bits of the third value in the digital input vector, and generates an analog optical signal 3988 that represents the third value of the digital input vector. The optical modulator 3970 converts the third 4-bit value of the digital input vector to a third analog optical signal. The optical modulator 3990 modulates the optical signal in the input waveguide 3050d according to the four bits of the fourth value in the digital input vector, and generates an analog optical signal 3998 that represents the fourth value of the digital input vector. The optical modulator 3990 converts the fourth 4-bit value of the digital input vector to a fourth analog optical signal. The analog optical signals 3968, 3978, 3988, and 3998 together form an analog optical input vector that can be provided to, e.g., the optoelectronic matrix multiplication unit 150 of
In this example, the laser unit 142 outputs four light signals on four optical waveguides 4050a, 4050b, 4050c, and 4050d. The light signal on the waveguide 4050a is modulated by the optical modulator 4080 according to the first 4-bit value of the digital input vector. The light signal on the waveguide 4050b is modulated by the optical modulator 4090 according to the second 4-bit value of the digital input vector. The light signal on the waveguide 4050c is modulated by the optical modulator 4100 according to the third 4-bit value of the digital input vector. The light signal on the waveguide 4050d is modulated by the optical modulator 4110 according to the fourth 4-bit value of the digital input vector.
The optical modulator 4080 includes two 1×2 port multi-mode interference couplers (MMI 1×2) 4012a and 4012b, two balanced arms 4014a and 4014b, and fifteen sub-modulators 4016_1 to 4016_15 in one arm (or fifteen sub-modulators in each arm). For example, each sub-modulator can include a phase shifter. The sub-modulator 4016_1 receives the LSB (B11) of the first value in the digital input vector. The two sub-modulators 4016_2 and 4016_3 receive the 2nd-bit (B12) of the first value in the digital input vector. The four sub-modulators 4016_4 to 4016_8 receive the 3rd-bit (B13) of the first value in the digital input vector. The eight sub-modulators 4016_9 to 4016_15 receive the MSB (B14) of the first value in the digital input vector.
Each of the sub-modulators 4016_1 to 4016_15 includes a waveguide segment, and the waveguide segments in the sub-modulators 4016_1 to 4016_15 are approximately the same length. The number of sub-modulators that receive the 2nd-bit (B12) is twice the number of sub-modulator that receives the LSB (B11). The number of sub-modulators that receive the 3rd-bit (B13) is twice the number of sub-modulator that receives the 2nd-bit (B12). The number of sub-modulators that receive the MSB (B14) is twice the number of sub-modulator that receives the 3rd-bit (B13).
The optical modulator 4090 includes fifteen sub-modulators in one arm (or fifteen sub-modulators in each arm). For example, each sub-modulator can include a phase shifter. One sub-modulator receives the LSB (B21) of the second value in the digital input vector. Two sub-modulators receive the 2nd-bit (B22) of the second value in the digital input vector. Four sub-modulators receive the 3rd-bit (B23) of the second value in the digital input vector. Eight sub-modulators receive the MSB (B24) of the second value in the digital input vector. The optical modulators 4100 and 4110 are configured in a manner similar to the optical modulators 4080 and 4090.
The optical modulator 4080 modulates the optical signal in the input waveguide 4050a according to the four bits of the first value in the digital input vector, and generates an analog optical signal 4018 that represents the first value of the digital input vector. The optical modulator 4080 converts the first 4-bit value of the digital input vector to a first analog optical signal. The optical modulator 4090 modulates the optical signal in the input waveguide 3050b according to the four bits of the second value in the digital input vector, and generates an analog optical signal 4020 that represents the second value of the digital input vector. The optical modulator 4090 converts the second 4-bit value of the digital input vector to a second analog optical signal.
The optical modulator 4100 modulates the optical signal in the input waveguide 3050c according to the four bits of the third value in the digital input vector, and generates an analog optical signal 4022 that represents the third value of the digital input vector. The optical modulator 4100 converts the third 4-bit value of the digital input vector to a third analog optical signal. The optical modulator 4110 modulates the optical signal in the input waveguide 3050d according to the four bits of the fourth value in the digital input vector, and generates an analog optical signal 4024 that represents the fourth value of the digital input vector. The optical modulator 4110 converts the fourth 4-bit value of the digital input vector to a fourth analog optical signal. The analog optical signals 4018, 4020, 4022, 4024 together form an analog optical input vector that can be provided to, e.g., the optoelectronic matrix multiplication unit 150 of
Various modifications can be made to the segmented design for the optical modulators described above.
The optical modulator 4170 includes seven sub-modulators 4172_1 to 4072_7 in one arm (or seven sub-modulators in each arm). For example, each sub-modulator can include a phase shifter. The sub-modulator 4172_1 receives the LSB (B11) of the first value in the digital input vector. The two sub-modulators 4172_2 and 4172_3 receive the 2nd-bit (B12) of the first value in the digital input vector. The two sub-modulators 4172_4 and 4172_5 receive the 3rd-bit (B13) of the first value in the digital input vector. The two sub-modulators 4172_6 and 4172_7 receive the MSB (B14) of the first value in the digital input vector.
Each of the sub-modulators 4172_1 to 4172_7 includes a waveguide segment. The waveguide segments in the sub-modulators 4172_1 to 4172_3 are approximately the same length. The waveguide segments in the sub-modulators 4172_4 and 4172_5 are twice as long as the waveguide segments in the sub-modulators 4172_2 and 4172_3. The waveguide segments in the sub-modulators 4172_6 and 4172_7 are twice as long as the waveguide segments in the sub-modulators 4172_4 and 4172_5. The optical modulators 4180, 4190, and 4200 are configured in a manner similar to the optical modulator 4170.
In the examples of
In some implementations, different optical modulators of a modulator array can have different segmentation designs. For example, in the modulator array 4060 of
Other device fabrication techniques can be combined with these techniques to further mitigate the effects of varying device characteristics. For example, dummy devices can be placed in certain locations to make the mask opening density more uniform. Those dummy devices can be left disconnected, so they have no effect on the optical modulation.
In some implementations, there can be an extra segment that is used for trim adjustments. For example, a 4-segment device can have a 5th segment that is used for trimming. The size of the trim segment can be the same size as the smallest segment, for example, and can be controlled separately (e.g., using a small applied voltage) to calibrate differences among different devices.
Driving circuits that include charge-pump bandwidth-enhancing circuits, similar to the circuit of
The segmented modulators of the modulator array 4060 (
For example, an input optical wave that enters the optical waveguide 5402a at a port A can propagate to a port B of the optical waveguide 5402a, or be coupled to a port C of the optical waveguide 5402b through the ring waveguide 5402c. The modulations imparted by the phase shifters 5404a and 5404b to the optical wave propagating in the ring waveguide 5402c modulates the amplitude of the optical wave at port B of the optical waveguide 5402a. The two phase shifters 5404a and 5404b effectively perform a 2-bit digital-to-analog conversion.
For example, the first phase shifter 5404a is driven by a first charge-pump bandwidth enhancing circuit 5406a, and the second phase shifter 5404b is driven by a second charge-pump bandwidth enhancing circuit 5406b. The charge-pump bandwidth enhancing circuits 5406a, 5406b are similar to the charge-pump bandwidth enhancing circuit shown in
For example, the first charge-pump bandwidth enhancing circuit 5406a is driven by a 2-bit DAC 5408a, and the second charge-pump bandwidth enhancing circuit 5406b is driven by a 2-bit DAC 5408b. The combination of the use of 2-bit DACs 5408a and 5408b and the two-segment ring resonator modulator 5400 allows a 4-bit digital input signal to be converted to an analog optical signal having 4-bit precision.
In some examples, there are n phase shifters that modulate the optical waves propagating in the ring waveguide 5402c, and the modulator is referred to as an n-segment ring resonator modulator. If each of the phase shifters is driven by a charge-pump bandwidth enhancing circuit that is in turn driven by a m-bit DAC (m>1), then the combination of the m-bit DACs and the n-segment ring resonator modulator can convert an m×n bit digital input signal to an analog optical signal having m×n bit precision. In general, the smaller the number m is, the faster the m-bit DAC can operate, and the faster the ring resonator modulator can be reconfigured.
In some implementations, the ring resonator modulator 5400 can be used in the modulator array 144 of
In some implementations, in a non-segmented design of an optical modulator driven by an electrical DAC signal, a single phase shifter in an interferometric modulator can also take advantage of signal conditioning for bandwidth enhancement. For example, if the interferometric modulator is a ring resonator (e.g., such as ring resonator 2222 of
Signal conditioning in a non-segmented design is potentially more complex than in a segmented design when there is pulse amplitude modulation (PAM) with more than two levels (i.e., for a digital value with more than 2 bits). In an interferometric modulator (e.g., the MZI modulator 3900 of
In some implementations, nonlinearity can occur in a modulator when mapping different modulator input values to different resulting amplitude reductions of the modulator. There are techniques for compensating for such nonlinearity, but the nonlinearity compensation process can be relatively slow and/or consume a relatively large amount of power. Techniques for faster and more power efficient pre-emphasis/de-emphasis signal conditioning with nonlinearity compensation can be achieved by combining different signal conditioning paths when forming the electrical signal that drives the single phase shifter in an interferometric modulator. A given number of bits for a series of input values (e.g., n bits), or a parallel n-bit value, can be used to control each of 2n single-level DACs. For example, the 2n single-level DACs can be included in the first MC subunit 132 (if the MZIs having segmented design are used for the modulator array 144), the second MC subunit 134 (if the MZIs having segmented design are used for the OMNI unit 150), or both (if the MZIs having segmented design are used for both the modulator array 144 and the OMM unit 150). Each single-level DAC is configured to provide an electrical signal having an amplitude for only one of the 2n bit values, and corresponding pre-emphasis and de-emphasis for that amplitude, that are appropriately conditioned for a given nonlinearity of the modulator at that amplitude. The electrical signal that provides a modulator input value is then selected from an appropriate one of the conditioned single-level DAC outputs.
For example, a first signal conditioning path from one of the single-level DACs providing a given modulator input value can be configured to provide an unconditioned electrical signal corresponding to the series of digital input values that drive that DAC, where for a given type of DAC (e.g., a current-steering DAC), the output can be scaled appropriately based on a known nonlinearity of the modulator at that modulator input value. A second signal conditioning path can be configured to provide a delayed, scaled, and/or inverted version of the unconditioned electrical signal (to provide pre-emphasis). A third signal conditioning path can be configured to provide a delayed, scaled, and/or inverted version of the unconditioned electrical signal (to provide de-emphasis). Any number of additional signal conditioning paths can be used, as needed, to provide the appropriate amount of pre-emphasis and/or de-emphasis when those different versions of the unconditioned electrical signal are added to the unconditioned electrical signal. The resulting electrical signals, which have been scaled to compensate for nonlinearity and have been conditioned to provide appropriate amounts of pre-emphasis and de-emphasis, can then be applied as modulator input values to result in equalized and bandwidth enhanced optical signals with different optical amplitudes evenly spaced over the 4 or more levels being used.
Referring back to
Additionally, the analog electronic unit 146 of the WDM ANN computation system 104 is further configured to demultiplex the multiple wavelengths and to generate a plurality of demultiplexed output voltages. For example, the analog electronic unit 146 can include a demultiplexer configured to demultiplex the three wavelengths contained in each of the 32 signals of the multi-wavelength optical output vector, and route the three single-wavelength optical output vectors to three banks of photodetectors coupled to three banks of transimpedance amplifiers.
Additionally, the ADC unit 160 of the WDM ANN computation system 104 includes banks of ADCs configured to convert the plurality of demultiplexed output voltages of the analog electronic unit 146. Each of the banks corresponds to one of the multiple wavelengths, and generates respective digitized demultiplexed optical outputs. For example, the banks of ADCs can be coupled to the banks of transimpedance amplifiers of the analog electronic unit 146.
The controller 110 can implement a method analogous to the process 200 but expanded to support the multi-wavelength operation. For example, the method can include the steps of obtaining, from the ADC unit 160, a plurality of digitized demultiplexed optical outputs, the plurality of digitized demultiplexed optical outputs forming a plurality of first digital output vectors, in which each of the plurality of first digital output vectors corresponds to one of the plurality of wavelengths; performing a nonlinear transformation on each of the plurality of first digital output vectors to generate a plurality of transformed first digital output vectors; and storing, in the memory unit, the plurality of transformed first digital output vectors.
In some examples, the ANN can be specifically designed, and the digital input vectors can be specifically formed such that the multi-wavelength optical output vector can be detected without demultiplexing. In such examples, the analog electronic unit 146 can be a wavelength-insensitive detection unit that does not demultiplex the multiple wavelengths of the multi-wavelength optical output vector. As such, each of the photodetectors of the analog electronic unit 146 effectively sums the multiple wavelengths of an optical signal into a single photocurrent, and each of the voltages output by the analog electronic unit 146 corresponds to an element-by-element sum of the matrix multiplication results of the multiple digital input vectors.
So far, the nonlinear transformations of the weighted sums performed as part of the ANN computation were performed in the digital domain by the controller 110. In some examples, the nonlinear transformations can be computationally intensive or power hungry, add significantly to the complexity of the controller 110, or otherwise limit the performance of the ANN computation system 100 or 104 in terms of throughput or power efficiency. As such, in some implementations of the ANN computation system 100 or 104, the nonlinear transformation can be performed in the analog domain through analog electronics.
In some implementations, the analog electronic unit 146 is configured to apply a nonlinear transfer function, and output transformed output voltages to the ADC unit 160. As the ADC unit 160 receives voltages that have been nonlinearly transformed by the analog electronic unit 146, the controller 110 can obtain, from the ADC unit 160, transformed digitized output voltages corresponding to the transformed output voltages. Because the digitized output voltages obtained from the ADC unit 160 have already been nonlinearly transformed (“activated”), the nonlinear transformation step by the controller 110 can be omitted, reducing the computation burden by the controller 110. The first transformed voltages obtained directly from the ADC unit 160 can then be stored as the first transformed digital output vector in the memory unit 120.
The analog electronic nonlinear transformation performed by the analog electronic unit 146 can be implemented in various ways. For example, high-gain amplifiers in feedback configuration, comparators with adjustable reference voltage, nonlinear IV (current—voltage) characteristics of a diode, breakdown behavior of a diode, nonlinear CV (capacitance—voltage) characteristics of a variable capacitor, or nonlinear IV characteristics of a variable resistor can be used.
Use of the analog nonlinear transformation can improve the performance, such as throughput or power efficiency, of the ANN computation system 104 by reducing a step to be performed in the digital domain. The moving of the nonlinear transformation step out of the digital domain can allow additional flexibility and improvements in the operation of the ANN computation systems. For example, in a recurrent neural network, the output of the OMNI unit 150 is activated, and recirculated back to the input of the OMM unit 150. In the examples in which the activation is performed by the controller 110 in the ANN computation system 100, it is necessary to digitize the output voltages of the detection unit 146 at every pass through the OMNI unit 150. In the examples in which the activation is performed prior to digitization by the ADC unit 160, it is possible to reduce the number of ADC conversions needed in performing recurrent neural network computations.
In some implementations, the analog nonlinear transformation can be performed by the ADC unit 160. For example, a nonlinear ADC unit can be a linear ADC unit with a nonlinear lookup table that maps the linear digitized outputs of the linear ADC unit into desired nonlinearly transformed digitized outputs.
Some implementations of the optoelectronic processor 140 of the ANN computation system 100 in
The optoelectronic computing system can produce a computational result using different types of operations that are each performed on signals (e.g., electrical signals or optical signals) for which the underlying physics of the operation is most suitable (e.g., in terms of energy consumption and/or speed). For example, copying can be performed using optical power splitting, summation can be performed using electrical current-based summation, and multiplication can be performed using optical amplitude modulation. An example of a computation that can be performed using these three types of operations is multiplying a vector by a matrix (e.g., as employed by artificial neural network computations). A variety of other computations can be performed using these operations, which represent a set of general linear operations from which a variety of computations can be performed, including but not limited to: vector-vector dot products, vector-vector element-wise multiplication, vector-scalar element wise multiplication, or matrix-matrix element-wise multiplication.
Referring to
In some implementations, the optoelectronic computing system 1800 is configured to perform a computation on an array of input values that are encoded on respective optical signals provided by the optical ports or sources 1802A, 1802B, etc. For example, for various machine learning applications based on neural networks, the computation can implement vector-matrix multiplication (or vector-by-matrix multiplication) where an input vector is multiplied by a matrix to yield an output vector as a result. The optical signals can represent elements of a vector, including possibly only a subset of selected elements of the vector. For example, for some neural network models, the size of a matrix used in the computation can be larger than the size of a matrix that can be loaded into a hardware system (e.g., an engine or co-processor of a larger system) that performs a vector-matrix multiplication portion of the computation. So, part of performing the computation can involve dividing the matrix and the vector into smaller segments that can be provided to the hardware system separately.
The modules shown in
by a matrix
to produce an output vector
For this vector-matrix multiplication
yA=MAxA+MBxB (Equ. 1)
yB=McxA+MDxB (Equ. 2)
The equations (1) and (2) can be broken down into separate steps that can be performed in the system 1800 using a set of basic operations: a copying operation, a multiplication operation, and a summation operation. In these equations, each element of the input vector appears twice, so there are two copying operations. There are also four multiplication operations, and there are two summation operations. The number of operations performed would be larger for systems that implement vector-matrix multiplication using a larger matrix, and the relative number of instances of each operation would be different using a matrix that is not square matrix in shape (i.e., with the number of rows being different from the number of columns).
In this example, the copying operations are performed by copying modules 1804A and 1804B. The elements of the input vector xA and xB are represented by values encoded on optical signals from the optical port/source 1802A and 1802B, respectively. Each of these values is used in both equations, so each value is copied to provide the resulting two copies to different respective multiplication modules. A value can be encoded in a particular time slot, for example, using optical wave that has been modulated to have a power from a set of multiple power levels, or having a duty cycle from a set of multiple duty cycles, as described in more detail below. A value is copied by copying the optical signal on which that value is encoded. The optical signal encoded with the value representing element xA is copied by copying module 1804A, and the optical signal encoded with the value representing element xB is copied by copying module 1804B. Each copying module can be implemented, for example, using an optical power splitter, such as a waveguide optical splitter that couples a guided mode in an input waveguide to each of two output waveguides over a Y-shaped splitter that gradually (e.g., adiabatically) splits the power, or a free-space beam splitter that uses a dielectric interface or thin film with one or more layers to transmit and reflect, respectively, two output beams from an input beam.
In this document, when we say that the optical signal encoded with the value representing element xA is copied by the copying module 1804A, we mean that multiple copies of signals that represent element xA are produced based on the input signal, not necessarily that the output signals of the copying module 1804A have the same amplitude as that of the input signal. For example, if the copying module 1804A splits the input signal power evenly between two output signals, then each of the two output signals will have a power that is equal to or less than 50% of the power of the input signal. The two output signals are copies of each other, while the amplitude of each output signal of the copying module 1804A is different from the amplitude of the input signal. Also, in some embodiments that have a group of multiple copying modules used for copying a given optical signal, or subset of optical signals, each individual copying module does not necessarily split power evenly among its generated copies, but the group of copying modules can be collectively configured to provide copies that have substantially equal power to the inputs of downstream modules (e.g., downstream multiplication modules).
In this example, the multiplication operations are performed by four multiplication modules 1806A, 1806B, 1806C, and 1806D. For each copy of one of the optical signals, one of the multiplication modules multiplies that copy of the optical signal by a matrix element value, which can be performed using optical amplitude modulation. For example, the multiplication module 1806A multiplies the input vector element xA by the matrix element MA. The value of the vector element xA can be encoded on optical signal, and the value of the matrix element MA can be encoded as an amplitude modulation level of an optical amplitude modulator.
The optical signal encoded with the vector element xA can be encoded using different forms of amplitude modulation. The amplitude of the optical signal can correspond to a particular instantaneous power level PA of a physical optical wave within a particular time slot, or can correspond to a particular energy EA of a physical optical wave over a particular time slot (where the power integrated over time yields total energy). For example, the power of a laser source can be modulated to have a particular power level from a predetermined set of multiple power levels. In some implementations, it can be useful to operate electronic circuitry near an optimized operation point, so instead of varying the power over many possible power levels, an optimized “on” power level is used with the signal being modulated to be “on” and “off” (at zero power) for particular fractions of a time slot. The fraction of time that the power is at the “on” level corresponds to a particular energy level. Either of these particular values of power or energy can be mapped to a particular value of the element xA (using a linear or nonlinear mapping relationship). The actual integration over time, to yield a particular total energy level, can occur downstream in the system 1800 after signals are in the electrical domain, as described in more detail below.
Additionally, the term “amplitude” can refer to the magnitude of the signal represented by the instantaneous or integrated power in the optical wave, or can also equivalently refer to the “electromagnetic field amplitude” of the optical wave. This is because the electromagnetic field amplitude has a well-defined relationship to the signal amplitude (e.g., by integrating an electromagnetic field intensity, which is proportional to the square of the electromagnetic field amplitude, over a transverse size of a guided mode or free-space beam to yield the instantaneous power). This leads to a relationship between modulation values, since a modulator that modulates the electromagnetic field amplitude by a particular value √{square root over (M)} can also be considered as modulating the power-based signal amplitude by a corresponding value M (since the optical power is proportional to the square of the electromagnetic field amplitude).
The optical amplitude modulator used by the multiplication module to encode the matrix element MA can operate by changing the amplitude of the optical signal (i.e., the power in the optical signal) using any of a variety of physical interactions. For example, the modulator can include a ring resonator, an electro-absorption modulator, a thermal electro-optical modulator, or a Mach-Zehnder Interferometer (MZI) modulator. In some techniques a fraction of the power is absorbed as part of the physical interaction, and in other techniques the power is diverted using a physical interaction that modifies another property of the optical wave other than its power, such as its polarization or phase, or modifies coupling of optical power between different optical structures (e.g., using tunable resonators). For optical amplitude modulators that operate using interference (e.g., destructive and/or constructive interference) among optical waves that have traveled over different paths, coherent light sources such as lasers can be used. For optical amplitude modulators that operate using absorption, either coherent or non-coherent or low-coherence light sources such as LEDs can be used.
In one example of a waveguide 1×2 optical amplitude modulator, a phase modulator is used to modulate the power in an optical wave by placing the phase modulator in one of multiple waveguides of the modulator. For example, the waveguide 1×2 optical amplitude modulator can split an optical wave guided by an input optical waveguide into first and second arms. The first arm includes a phase shifter that imparts a relative phase shift with respect to a phase delay of the second arm. The modulator then combines the optical waves from the first and second arms. In some embodiments, different values of the phase delay provide multiplication of the power in the optical wave guided by the input optical waveguide by a value between 0 to 1 through constructive or destructive interference. In some embodiments, the first and second arms are combined into each of two output waveguides, and a difference between photocurrents generated by respective photodetectors receiving light waves from the two output waveguides provides a signed multiplication result (e.g., multiplication by a value between −1 to 1), as described in more detail below (see
In this example, the summation operations are performed by two summation modules, with the summation module 1808, shown in
In embodiments in which the summation is performed in the electrical domain, the summation module 1808 can be implemented using: (1) two or more input conductors that each carries an input current whose amplitude represents a result of one of the multiplication modules, and (2) at least one output conductor that carries a current that is the sum of the input currents. For example, this occurs if the conductors are wires that meet at a junction. Such a relationship can be understood, for example (without being bound by theory), based on Kirchhoff's current law, which states that current flowing into a junction is equal to current flowing out of the junction. For these embodiments, the signals 1810A and 1810B provided to the summation module 1808 are input currents, which can be produced by photodetectors that are part of the multiplication modules that generate a respective photocurrent whose amplitude is proportional to the power in a received optical signal. The summation module 1808 then provides the output current isum. The instantaneous value of that output current, or the integrated value of that output current, can then be used to represent the quantitative value of the sum.
In embodiments in which the summation is performed in the optoelectronic domain, the summation module 1808 can be implemented using a photodetector (e.g., a photodiode) that receives the optical signals generated by different respective multiplication modules. For these embodiments, the signals 1810A and 1810B provided to the summation module 1808 are input optical signals that each comprise an optical wave whose power represents a result of one of the multiplication modules. The output current isum in this embodiment is the photocurrent generated by the photodetector. Since the wavelengths of the optical waves are different (e.g., different enough such that no significant constructive or destructive interference occurs between them), the photocurrent will be proportional to the sum of the powers of the received optical signals. The photocurrent is also substantially equal to the sum of the individual currents that would result for the individual detected optical powers detected by separate equivalent photodetectors. The wavelengths of the optical waves are different, but close enough to have substantially the same response by the photodetector (e.g., wavelengths within a substantially flat detection bandwidth of the photodetector). As mentioned above, summation in the electrical domain, using current summation, can enable a simpler system architecture by avoiding the need for multiple wavelengths.
and the matrix is
Each of the elements of the input vector is encoded on a different optical signal. Two different copying modules 1902a, 1902b (collectively referenced as 1902) perform an optical copying operation to split the computation over different paths (e.g., an “upper” path and a “lower” path, in which the term “upper” and “lower” refer to the relative positions of the paths in the figure). There are four multiplication modules 1904a, 1904b, 1904c, 1904d (collectively referenced as 1904) that each multiply by a different matrix element using optical amplitude modulation. At the output of each multiplication module 1904, there is an optical detection module 1906 (e.g., 1906a, 1906b, 1906c, 1906d) that converts an optical signal to an electrical signal in the form of an electrical current. Both upper paths of the different input vector elements (e.g., including the outputs of 1906a and 1906c) are combined using a summation module 1908a, and both lower paths of the different input vector elements (including the outputs of 1906b and 1906d) are combined using a summation module 1908b. The summation modules 1908a and 1908b (collectively referenced as 1908) perform summation in the electrical domain. So, each of the elements of output vector is encoded on a different electrical signal. As shown in
M11v1+M12v2
M21v1+M22v2
At different portions of the system, different optical powers can represent the same value. For example, the copying module 1902a receives an input signal on an input waveguide 1914 and provides output signals on output waveguides 1916a and 1916b. The optical signal on the output waveguide 1916a or 1916b that represents the value v1 has an amplitude that is approximately half of the amplitude of the optical signal on the input waveguide 1914 that represents the value v1.
In some implementations, if a copying module performs an optical copying operation to split the computation over three paths, then the optical signal on the output waveguide of the optical splitter that represents a particular value has an amplitude that is approximately one-third of the amplitude of the optical signal on the input waveguide of the optical splitter that represents the particular value. Similarly, if a copying module performs an optical copying operation to split the computation over four paths, then the optical signal on the output waveguide of the optical splitter that represents a particular value has an amplitude that is approximately one-fourth of the amplitude of the optical signal on the input waveguide of the optical splitter that represents the particular value, and so forth.
In some implementations, a photonic integrated circuit includes different types of copying modules, e.g., a first copying module that performs an optical copying operation to split the computation over two paths, a second copying module that performs an optical copying operation to split the computation over three paths, a third copying module that performs an optical copying operation to split the computation over four paths, and a fourth copying module that performs an optical copying operation to split the computation over eight paths. Signals derived from the outputs of the first, second, third, and fourth copying modules are scaled before they are combined.
For example, suppose vout1 is a value of a vector resulting from a vector-matrix multiplication using a 2×2-element matrix, in which a 1-to-2 splitter is used in the optical copying operation, and vout2 is a value of a vector resulting from a vector-matrix multiplication using a 4×4-element matrix, in which a 1-to-4 splitter is used in the optical copying operation. If the photonic integrated circuit is configured such that vout1 is combined with vout2, then vout2 is scaled to twice its value before being combined with vout1.
The system configuration 1900 can be implemented using any of a variety of optoelectronic technologies. In some implementations, there is a common substrate (e.g., a semiconductor such as silicon), which can support both integrated optics components and electronic components. The optical paths can be implemented in waveguide structures that have a material with a higher optical index surrounded by a material with a lower optical index defining a waveguide for propagating an optical wave that carries an optical signal. The electrical paths can be implemented by a conducting material for propagating an electrical current that carries an electrical signal. In
In this document, a figure can show an optical waveguide crossing an electrical signal line, it is understood that the optical waveguide does not intersect the electrical signal line. The electrical signal line and the optical waveguide can be disposed at different layers of the device.
In this document, when a figure shows two optical waveguides crossing each other, whether the two optical waveguides are actually optically coupled to each other will be clear from the description. For example, two waveguides that appear to cross each other from a top view of the device can be implemented in different layers and thus not intersect with each other. For example, in
The system configurations shown in
and the matrix is
For example, the input vector elements v1 to vn are provided by n waveguides, and each input vector element is processed by one or more copying modules to provide m copies of the input vector element to m respective paths. There are m×n multiplication modules that each multiply by a different matrix element using optical amplitude modulation to produce an electrical or optical signal representing Mij·vj (i=1 . . . m, j=1 . . . n). The signals representing Mij·v1 (j=1 . . . n) are combined using an i-th summation module (i=1 . . . m) to produce the following results for the m paths, respectively.
M11v1+M12v2+ . . . +M1nvn
M21v1+M22v2+ . . . +M2nvn
. . .
Mm1v1+Mm2v2+ . . . +Mmnvn
Since optical amplitude modulation is able to reduce the power in an optical signal from its full value to a lower value, down to zero (or near zero) power, multiplication by any value between 0 and 1 can be implemented. However, some computations can call for multiplication by values greater than 1 and/or multiplication by signed (positive or negative) values. First, for extending the range to 0 to Mmax (where Mmax>1), the original modulation of the optical signals can include an explicit or implicit scaling of an original vector element amplitude by Mmax (or equivalently, scaling the value mapped to a particular vector element amplitude in a linear mapping by 1/Mmax) such that the range 0 to 1 for matrix element amplitudes corresponds quantitatively in the computation to the range 0 to Mmax. Second, for extending the positive range 0 to Mmax for matrix element values to a signed range −Mmax to Mmax, a symmetric differential configuration can be used, as described in more detail below. Similarly, a symmetric differential configuration can also be used to extend a positive range for the values encoded on the various signals to a signed range of values.
V1=V1+−V1−
where the signed value V1 monotonically increases between −Vmax and Vmax as the unsigned main value V1+ monotonically increases from 0 to Vmax and its paired anti-symmetric value V1− monotonically decreases from Vmax to 0. There are various techniques that can be used for implementing the symmetric differential configuration of
In
In some embodiments of a detector that uses a TIA to convert a photocurrent to a voltage, techniques are used to mitigate the effects of the internal capacitance of the photodiodes used as the photodetectors 2012 and 2016. The internal capacitance Cd of a photodiode can be modeled as a capacitor in parallel with an ideal photodiode. One consequence is that the capacitor acts as a short circuit when the current changes quickly, which could limit the bandwidth of the detector. Generally, the time constant τ associated with changes resulting from the input current 2020 to the op-amp 2030 should be kept small to avoid a significant bandwidth limitation. Without mitigation, this time constant would be approximately equal to the capacitance Cd multiplied by the input resistance Rin, of the op-amp 2030, which is approximately the feedback resistance Rf due to the resistive element 2028 divided by the op-amp gain A, yielding: τ=CdRf/A. The effect could be particularly detrimental in a system that has a large number of photodiodes in parallel with each other as in some of the systems described herein, since capacitances in parallel with each other add to a large effective capacitance. However, achieving a large op-amp gain A to reduce this time constant takes a significant amount of power in a system that uses a large number of such TIA circuits. To mitigate the bandwidth limitations without requiring as much power, a voltage follower circuit (also called a unity-gain amplifier or a buffer amplifier) can be placed in parallel with each photodiode by connecting a given voltage follower to the input and output terminals of a given photodiode. The voltage follower provides current as needed to maintain an approximately equal voltage at its input and output terminals, which has the effect of preventing the internal capacitance of the photodiode from acting as a short circuit when the current changes quickly, achieving a near-zero time constant τ. Use of the voltage followers enables the op-amp gain A in the TIA circuits to be relatively low (e.g., lower than 1000 or lower than 100), reducing the power requirements of the system.
In
M11V=M11+V−M11−V
where the signed value M11 monotonically increases between −Mmax and Mmax as the unsigned main value M11+ monotonically increases from 0 to Mmax and its paired anti-symmetric value M11− monotonically decreases from Mmax to 0.
The system configuration 2110 also includes other modules arranged as shown in
M11V1M12V2=(M11+−M11−)(V1+−V1−)(M12+−M12−)(V2+−V2−)
M21V1M22V2=(M21+−M21−)(V1+−V1−)(M22+−M22−)(V2+−V2−)
In this document, when a figure shows two electrical signal lines crossing each other, whether the two electrical signal lines are electrically coupled to each other will be clear from the description. For example, the signal line carrying the M21+V1+ signal is not electrically coupled to the signal line carrying the M11+V1− signal or the signal line carrying the M11−V1− signal.
The system configuration shown in
There are various techniques that can be used for implementing the symmetric differential configuration of
The 1×2 optical amplitude modulator 2200 includes a 2×2 coupler 2206 that combines the optical waves from first and second input paths using optical interference or optical coupling in a particular manner to divert power into first and second output paths in different ratios, depending on the phase shift. For example, in a free-space interferometer, a phase shift of 0 degrees causes substantially all of the input power that was split between the two paths to constructively interfere to exit from one output path of a beam splitter implementing the coupler 2206, and a phase shift of 180 degrees causes substantially all of the input power that was split between the two paths constructively interfere to exit from the other output path of the beam splitter implementing the coupler 2206. In a waveguide interferometer, a phase shift of 0 degrees causes substantially all of the input power that was split between the two paths to couple to one output waveguide (e.g., 2208a) of the coupler 2206, and a phase shift of 180 degrees causes substantially all of the input power that was split between the two paths to couple to the other output waveguide (e.g., 2208b) of the coupler 2206. Phase shifts between 0 degrees and 180 degrees can then provide multiplication of the power in an optical wave (and the value encoded on the optical wave) by a value between 0 and 1 through partial constructive or destructive interference, or partial waveguide coupling. Multiplication by any value between 0 to 1 can then be mapped to multiplication by any value between 0 to Mmax as described above.
Additionally, the relationship between the power in the two optical waves emitted from the modulator 2200 follows that of the main and anti-symmetric pairs described above. When the amplitude of the optical power of one signal increases, the amplitude of the optical power of the other signal decreases, so a difference between detected photocurrents can yield a signed vector element, or multiplication by a signed matrix element, as described herein. For example, the pair of related optical signals can be provided from the two output ports of the modulator 2200 such that a difference between amplitudes of the related optical signals corresponds to a result of multiplying an input value by a signed matrix element value.
Other techniques can be used to construct 1×2 optical amplitude modulators for implementing the multiplication modules 1904, and/or for providing pairs of optical signals that are related as main and anti-symmetric pairs.
Since the time scale over which the optical power circulates around the ring resonator 2222 is small compared to the time scale of the amplitude modulation of the optical signals, an anti-symmetric power relationship is quickly established between the two output ports, such that the optical wave detected by the photodetector 2212 and the optical wave detected by the photodetector 2214 form main and anti-symmetric pairs. The resonance wavelength of the ring resonator 2222 can be tuned to monotonically decrease/increase the main/anti-symmetric signals to achieve a signed result, as described above. When the ring is completely off-resonance all of the power exits over the path 2224 out of the first output port, and when it is completely on-resonance, with certain other parameters (e.g., quality factor, and coupling coefficient) appropriately tuned, all of the power exits over the path 2228 out of the second output port. In particular, to achieve complete power transfer, the coupling coefficient characterizing the coupling efficiency between the waveguide and the ring resonator should be matched. In some embodiments, it is useful to have a relatively shallow tuning curve, which can be achieved by reducing the quality factor of the ring resonator 2222 (e.g., by increasing the loss) and correspondingly increasing the coupling coefficients into and out of the ring. A shallow tuning curve provides less sensitivity of the amplitude to the resonance wavelength. Techniques such as temperature control can also be used for tuning and/or stability of the resonance wavelength.
The different VMM subsystems are arranged so that the results of each submatrix are appropriately combined to yield results for the larger combined matrix (e.g., elements of a 128-element vector resulting from multiplication by a 128×128-element matrix). Each set of optical ports or sources 2402 provides a set of optical signals that represent different subsets of vector elements of a larger input vector. Copy modules 2404 are configured to copy all of the optical signals within a received set of optical signals encoded on optical waves guided in a set 2403 of 64 optical waveguides, and provide that set of optical signals to each of two different sets of optical waveguides, which in this example are a set 2405 A of 64 optical waveguides and a set 2405B of 64 optical waveguides. This copying operation can be performed, for example, by using an array of waveguide splitters, each splitter in the array copying one of the elements of the subset of input vector elements (e.g., a subset of 64 elements for each copy module 2404) by splitting an optical wave in the set 2403 of optical waveguides into a first corresponding optical wave in the set 2405A of optical waveguides and a second corresponding optical wave in the set 2405B of optical waveguides.
If multiple wavelengths are used in some embodiments (e.g., W wavelengths), the number of separate waveguides (and thus the number of separate ports or sources in 2402) can be reduced, for example, by a factor of 1/W. Each VMM subsystem device 2410 performs vector-matrix multiplication, providing its partial results as a set of electrical signals (for a subset of elements of the output vector), with corresponding partial result pairs from different devices 2410 being added together by the summation modules 2414 as shown in
In some implementations, vector-matrix multiplications using a desired matrix can be performed, recursively, by combining results from smaller submatrices, for any number of levels of recursion, ending by using the single element optical amplitude modulator at the root level of the recursion. At different levels of recursion the VMM subsystem device can be more compact (e.g., different data centers connected by long distance optical fiber networks at one level, different multi-chip devices connected by optical fibers within a data center at another level, different chips within a device connected by optical fibers at another level, and different sections of modules on the same chip connected by on-chip waveguides at another level).
In some implementations, a digital controller (not shown in the figure) is provided to control the operations of the data storage subsystem 2450, the hierarchical cache modules, various circuitry such as the digital-to-analog converters and analog-to-digital converters, the VMM subsystems 2410, and the optical sources 2402. For example, the digital controller is configured to execute program code to implement a neural network having several hidden layers. The digital controller iteratively performs matrix processing associated with various layers of the neural network. The digital controller performs a first iteration of matrix processing by retrieving first matrix data from the data storage subsystem 2450 and setting the modulation levels of the optical amplitude modulators in the VMM subsystems 2410 based on the retrieved data, in which the first matrix data represent coefficients of a first layer of the neural network. The digital controller retrieves a set of input data from the data storage subsystem and sets the modulation levels for the optical sources 2402 to produce a set of optical input signals that represent elements of a first input vector.
The VMM subsystems 2410 perform matrix processing based on the first input vector and the first matrix data, representing the processing of signals by the first layer of the neural network. After the auxiliary processing subsystem 2450 has produced a first set of result data 2462, the digital controller performs a second iteration of matrix processing by retrieving second matrix data from the data storage subsystem that represent coefficients of a second layer of the neutral network, and setting the modulation levels of the optical amplitude modulators in the VMM subsystems 2410 based on the second matrix data. The first set of result data 2462 is used as a second input vector to set the modulation levels for the optical sources 2402. The VMM subsystems 2410 perform matrix processing based on the second input vector and the second matrix data, representing the processing of signals by the second layer of the neural network, and so forth. At the last iteration, the output of the processing of signals by the last layer of the neural network is produced.
In some implementations, when performing computations associated with hidden layers of a neural network, the result data 2462 are not sent to the data storage subsystem 2450, but are used by the digital controller to directly control digital-to-analog converters that produce control signals for setting the modulation levels of the optical amplitude modulators in the VMM subsystems 2410. This reduces the time needed for storing data to and accessing data from the data storage subsystem 2450.
Other processing techniques can be incorporated into other examples of system configurations. For example, various techniques used with other kinds of vector-matrix multiplication subsystems (e.g., subsystems using optical interference without the electrical summation or signed multiplication described herein) can be incorporated into some system configurations, such as some of the techniques described in U.S. Patent Publication No. 2017/0351293, incorporated herein by reference.
Referring to
In some examples, the OMM unit 3220 includes MZIs without the segmented design, and the second DAC subunit 3206 can be similar to the second DAC subunit 134 of
In some examples, the OMM unit 3220 includes MZIs having the segmented design, and the second DAC subunit 3206 can include 1-bit DACs similar to the first DAC subunit 3204. The second DAC subunit 3206 can also include the charge pump driving circuits and perform pre-emphasis and de-emphasis to enhance the bandwidth, similar to the first DAC subunit 3204.
An optoelectronic processor 3210 includes a light source 3230, which can be similar to the laser unit 142 of
Referring to
and multiplies the input vector with a matrix
to produce an output vector
The optoelectronic matrix multiplication unit 3220 includes m optical paths 1803_1, 1803_2, . . . , 1803_m (collectively referenced as 1803) that carry the optical signals representing the input vector. A copying module 1804_1 provides copies of the input optical signal v1 to multiplication modules 1806_11, 1806_21, . . . , 1806_m1. A copying module 1804_2 provides copies of the input optical signal v2 to multiplication modules 1806_12, 1806_22, . . . , 1806_m2. A copying module 1804_n provides copies of the input optical signal vn to multiplication modules 1806_1n, 1806_2n, . . . , 1806_mn.
As discussed above, the amplitudes of the copies of the optical signal v1 provided by the copying module 1804_1 are the same (or substantially the same) relative to one another, but different from that of the optical signal v1 provided by the modulator array 3208. For example, if the copying module 1804_1 splits the signal power of v1 provided by the modulator array 3208 evenly among m signals, then each of the m signals will have a power that is equal to or less than 1/m of the power of v1 provided by the modulator array 3208.
A multiplication module 1806_11 multiplies the input signal v1 with a matrix element M11 to produce M11·v1. A multiplication module 1806_21 multiplies the input signal v1 with a matrix element M21 to produce M21·v1. A multiplication module 1806_m1 multiplies the input signal v1 with a matrix element Mm1 to produce Mm1·v1. A multiplication module 1806_12 multiplies the input signal v2 with a matrix element M12 to produce M12·v2. A multiplication module 1806_22 multiplies the input signal v2 with a matrix element M22 to produce M22·v2. A multiplication module 1806_m2 multiplies the input signal v2 with a matrix element Mm2 to produce Mm2·v2. A multiplication module 1806_1n multiplies the input signal vn with a matrix element M1 to produce M1n·vn. A multiplication module 1806_2n multiplies the input signal vn with a matrix element M2n to produce M2n·vn. A multiplication module 1806_mn multiplies the input signal vn with a matrix element Mmn to produce Mmn·vn, and so forth.
The second DAC subunit 3206 generates control signals based on the values of the matrix elements, and sends the control signals to the multiplication modules 1806 to enable the multiplication modules 1806 to multiply the values of the input vector elements with the values of the matrix elements, e.g., by using optical amplitude modulation. For example, the multiplication module 1806_11 can include an optical amplitude modulator, and multiplying the input vector element v1 by the matrix element M11 can be achieved by encoding the value of the matrix element M11 as an amplitude modulation level applied to the input optical signal representing the input vector element v1.
A summation module 1808_1 receives the outputs of the multiplication modules 1806_11, 1806_12, . . . , 1806_1n, and generates a sum y1 equal to M11v1+M12v2+ . . . + M1nvn. A summation module 1808_2 receives the outputs of the multiplication modules 1806_21, 1806_22, . . . , 1806_2n, and generates a sum y2 equal to M21v1+M22v2 . . . M2nvn. A summation module 1808_m receives the outputs of the multiplication modules 1806_m1, 1806_m2, . . . , 1806_mn, and generates a sum ym equal to Mm1v1+Mm2v2+ . . . +Mmnvn.
In the system 3200, the output of the optoelectronic matrix multiplication unit 3220 is provided to the ADC unit 160 without passing through a detection unit 146 as is the case in the system 100 of
At 3310, an artificial neural network (ANN) computation request comprising an input dataset and a first plurality of neural network weights is received. The input dataset includes a first digital input vector. The first digital input vector is a subset of the input dataset. For example, it can be a sub-region of an image. The ANN computation request can be generated by various entities, such as the computer 102 of
At 3320, the input dataset and the first plurality of neural network weights are stored. The controller 110 can store the input dataset and the first plurality of neural network weights in the memory unit 120. Storing of the input dataset and the first plurality of neural network weights in the memory unit 120 can allow flexibilities in the operation of the ANN computation system 3200 that, for example, can improve the overall performance of the system. For example, the input dataset can be divided into digital input vectors of a set size and format by retrieving desired portions of the input dataset from the memory unit 120. Different portions of the input dataset can be processed in various order, or be shuffled, to allow various types of ANN computations to be performed. For example, shuffling can allow matrix multiplication by block matrix multiplication technique in cases where the input and output matrix sizes are different. As another example, storing of the input dataset and the first plurality of neural network weights in the memory unit 120 can allow queuing of multiple ANN computation requests by the ANN computation system 3200, which can allow the system 3200 to sustain operation at its full speed without periods of inactivity.
In some implementations, the memory unit 120 can include a first memory subunit and a second memory subunit. The first memory subunit can be a memory dedicated to storing of the input dataset and the digital input vectors, and can have an operating speed sufficient to support the modulation rate of the modulator array 3208. The second memory subunit can be a memory dedicated to storing of the neural network weights, and can have an operation speed sufficient to support the reconfiguration rate of the OMNI unit 3220. In some implementations, the first memory subunit can be implemented using SRAM and the second memory subunit can be implemented using DRAM. In some implementations, the first and second memory subunits can be implemented using DRAM. In some implementations, the first memory unit can be implemented as a part of or as a cache of the controller 110. In some implementations, the first and second memory subunits can be implemented by a single physical memory device as different address spaces. In some implementations, the input dataset can be stored in the first memory subunit, and the first plurality of neural network weights can be stored in the second memory subunit.
At 3330, a first plurality of modulator control signals is generated based on the first digital input vector and a first plurality of weight control signals is generated based on the first plurality of neural network weights. The controller 110 can send a first DAC control signal to the modulator control unit 3202 for generating the first plurality of modulator control signals. The first DAC subunit 3204 generates the first plurality of modulator control signals based on the first DAC control signal, and the modulator array 3208 generates the optical input vector representing the first digital input vector.
The first DAC control signal can include multiple digital values to be converted by the 1-bit DACs in the first DAC subunit 3204 into the first plurality of modulator control signals. The multiple digital values are generally in correspondence with the first digital input vector, and can be related through various mathematical relationships or look-up tables. For example, the multiple digital values can be linearly proportional to the values of the elements of the first digital input vector. As another example, the multiple digital values can be related to the elements of the first digital input vector through a look-up table configured to maintain a linear relationship between the digital input vector and the optical input vector generated by the modulator array 3208.
The controller 110 can send a second DAC control signal to the modulator control unit 3202 for generating the first plurality of weight control signals. The second DAC subunit 3206 generates the first plurality of weight control signals based on the second DAC control signal, and the optoelectronic matrix multiplication unit 3220 is reconfigured according to the first plurality of weight control signals, implementing a matrix corresponding to the first plurality of neural network weights.
The second DAC control signal can include multiple digital values to be converted by the second DAC subunit 3206 into the first plurality of weight control signals. The multiple digital values are generally in correspondence with the first plurality of neural network weights, and can be related through various mathematical relationships or look-up tables. For example, the multiple digital values can be linearly proportional to the first plurality of neural network weights. As another example, the multiple digital values can be calculated by performing various mathematical operations on the first plurality of neural network weights to generate weight control signals that can configure the optoelectronic matrix multiplication unit 3220 to perform a matrix multiplication corresponding to the first plurality of neural network weights.
At 3340, a first plurality of digitized outputs corresponding to the electronic output vector of the optoelectronic matrix multiplication unit 3220 is obtained. The optical input vector generated by the modulator array 3208 is processed by the optoelectronic matrix multiplication unit 3220 and transformed into an electrical output vector. The electrical output vector is converted into digitized values by the ADC unit 160. The controller 110 can, for example, send a conversion request to the ADC unit 160 to begin a conversion of the voltages output by the optoelectronic matrix multiplication unit 3220 into digitized outputs. Once the conversion is complete, the ADC unit 160 can send the conversion result to the controller 110. Alternatively, the controller 110 can retrieve the conversion result from the ADC unit 160. The controller 110 can form, from the digitized outputs, a digital output vector that corresponds to the result of the matrix multiplication of the input digital vector. For example, the digitized outputs can be organized, or concatenated, to have a vector format.
In some implementations, the ADC unit 160 can be set or controlled to perform an ADC conversion based on a DAC control signal issued to the modulator control unit 3202 by the controller 110. For example, the ADC conversion can be set to begin at a preset time following the generation of the modulation control signal by the modulator control unit 3202. Such control of the ADC conversion can simplify the operation of the controller 110 and reduce the number of necessary control operations.
At 3350, a nonlinear transformation is performed on the first digital output vector to generate a first transformed digital output vector. A node, or an artificial neuron, of an ANN operates by first performing a weighted sum of the signals received from nodes of a previous layer, then performing a nonlinear transformation (“activation”) of the weighted sum to generate an output. Various types of ANN can implement various types of differentiable, nonlinear transformations. Examples of nonlinear transformation functions include a rectified linear unit (RELU) function, a Sigmoid function, a hyperbolic tangent function, an X2 function, and a |X| function. Such nonlinear transformations are performed on the first digital output by the controller 110 to generate the first transformed digital output vector. In some implementations, the nonlinear transformations can be performed by a specialized digital integrated circuitry within the controller 110. For example, the controller 110 can include one or more modules or circuit blocks that are specifically adapted to accelerate the computation of one or more types of nonlinear transformations.
At 3360, the first transformed digital output vector is stored. The controller 110 can store the first transformed digital output vector in the memory unit 120. In cases where the input dataset is divided into multiple digital input vectors, the first transformed digital output vector corresponds to a result of the ANN computation of a portion of the input dataset, such as the first digital input vector. As such, storing of the first transformed digital output vector allows the ANN computation system 3200 to perform and store additional computations on other digital input vectors of the input dataset to later be aggregated into a single ANN output.
At 3370, an artificial neural network output generated based on the first transformed digital output vector is output. The controller 110 generates an ANN output, which is a result of processing the input dataset through the ANN defined by the first plurality of neural network weights. In cases where the input dataset is divided into multiple digital input vectors, the generated ANN output is an aggregated output that includes the first transformed digital output, but can further include additional transformed digital outputs that correspond to other portions of the input dataset. Once the ANN output is generated, the generated output is sent to a computer, such as the computer 102, that originated the ANN computation request.
In some implementations, the controller 110, the memory unit 120, the modulator control unit 3202, the ADC unit 160, and a microprocessor can be monolithically integrated on a semiconductor die. In some implementations, the controller 110, the memory unit 120, the modulator control unit 3202, the ADC unit 160, the microprocessor, and the system main memory can be integrated as system-on-chip. In such examples, two or more of the steps in the process 3300 can be performed by the various modules in the monolithic integrated circuit or system-on-chip.
Various performance metrics can be defined for the ANN computation system 3200 implementing the method 3300. Defining performance metrics can allow a comparison of performance of the ANN computation system 3200 that implements the optoelectronic processor 3210 with other systems for ANN computation that instead implement electronic matrix multiplication units. In one aspect, the rate at which an ANN computation can be performed can be indicated in part by a first loop period defined as a time elapsed between the step 3320 of storing, in the memory unit, the input dataset and the first plurality of neural network weights, and the step 3360 of storing, in the memory unit, the first transformed digital output vector. This first loop period therefore includes the time taken in converting the electrical signals into optical signals (e.g., step 3330), and performing the matrix multiplication in the optical and electrical domains (e.g., step 3340). Steps 3320 and 3360 both involve storing of data into the memory unit 120, which are steps shared between the ANN computation system 3200 and conventional ANN computation systems without the optoelectronic processor 3210. As such, the first loop period measuring the memory-to-memory transaction time can allow a realistic or fair comparison of ANN computation throughput to be made between the ANN computation system 3200 and ANN computation systems without the optoelectronic processor 3210, such as systems implementing electronic matrix multiplication units.
Due to the rate at which the optical input vectors can be generated by the modulator array 3208 (e.g., at 25 GHz) and the processing rate of the optoelectronic matrix multiplication unit 3220 (e.g., >25 GHz), the first loop period of the ANN computation system 3200 for performing a single ANN computation of a single digital input vector can approach the reciprocal of the speed of the modulator array 3208, e.g., 40 ps. After accounting for latencies associated with the signal generation by the modulator control unit 3202 and the ADC conversion by the ADC unit 160, the first loop period can, for example, be less than or equal to 100 ps, less than or equal to 200 ps, less than or equal to 500 ps, less than or equal to 1 ns, less than or equal to 2 ns, less than or equal to 5 ns, or less than or equal to 10 ns.
As a comparison, execution time of a multiplication of an M×1 vector and an M×M matrix by an electronic matrix multiplication unit is typically proportional to M2−1 processor clock cycles. For M=32, such multiplication would take approximately 1024 cycles, which at 3 GHz clock speed results in an execution time exceeding 300 ns, which is orders of magnitude slower than the first loop period of the ANN computation system 3200.
In some implementations, the method 3300 further includes a step of generating a second plurality of modulator control signals based on the first transformed digital output vector. In some types of ANN computations, a single digital input vector can be repeatedly propagated through, or processed by, the same ANN. As previously discussed, an ANN that implements multi-pass processing can be referred to as a recurrent neural network (RNN). A RNN is a neural network in which the output of the network during a (k)th pass through the neural network is recirculated back to the input of the neural network and used as the input during the (k+1)th pass. RNNs can have various applications in pattern recognition tasks, such as speech or handwriting recognition. Once the second plurality of modulator control signals are generated, the method 3300 can proceed from step 3340 through step 3360 to complete a second pass of the first digital input vector through the ANN. In general, the recirculation of the transformed digital output to be the digital input vector can be repeated for a preset number of cycles depending of the characteristics of the RNN received in the ANN computation request.
In some implementations, the method 3300 further includes a step of generating a second plurality of weight control signals based on a second plurality of neural network weights. In some cases, the artificial neural network computation request further includes a second plurality of neural network weights. As previously discussed, in general, an ANN has one or more hidden layers in addition to the input and output layers. For ANN with two hidden layers, the second plurality of neural network weights can correspond, for example, to the connectivity between the first layer of the ANN and the second layer of the ANN. To process the first digital input vector through the two hidden layers of the ANN, the first digital input vector can first be processed according to the method 3300 up to step 3360, at which the result of processing the first digital input vector through the first hidden layer of the ANN is stored in the memory unit 120. The controller 110 then reconfigures the optoelectronic matrix multiplication unit 3220 to perform the matrix multiplication corresponding to the second plurality of neural network weights associated with the second hidden layer of the ANN. Once the optoelectronic matrix multiplication unit 3220 is reconfigured, the method 3300 can generate the plurality of modulator control signals based on the first transformed digital output vector, which generates an updated optical input vector corresponding to the output of the first hidden layer. The updated optical input vector is then processed by the reconfigured optoelectronic matrix multiplication unit 3220 which corresponds to the second hidden layer of the ANN. In general, the described steps can be repeated until the digital input vector has been processed through all hidden layers of the ANN.
In some implementations of the optoelectronic matrix multiplication unit 3220, the reconfiguration rate of the optoelectronic matrix multiplication unit 3220 can be significantly slower than the modulation rate of the modulator array 3208. In such cases, the throughput of the ANN computation system 3200 can be adversely impacted by the amount of time spent in reconfiguring the optoelectronic matrix multiplication unit 3220 during which ANN computations cannot be performed. To mitigate the impact of the relatively slow reconfiguration time of the optoelectronic matrix multiplication unit 3220, batch processing techniques can be utilized in which two or more digital input vectors are propagated through the optoelectronic matrix multiplication unit 3220 without a configuration change to amortize the reconfiguration time over a larger number of digital input vectors.
Referring to
The multiple wavelengths can preferably be separated by a wavelength spacing that is sufficiently large to allow easy multiplexing and demultiplexing onto a common propagation channel. For example, the wavelength spacing greater than 0.5 nm, 1.0 nm, 2.0 nm, 3.0 nm, or 5.0 nm can allow simple multiplexing and demultiplexing. On the other hand, the range between the shortest wavelength and the longest wavelength of the multiple wavelengths (“WDM bandwidth”) can preferably be sufficiently small such that the characteristics or performance of the optoelectronic matrix multiplication unit 3520 remain substantially the same across the multiple wavelengths. Optical components are typically dispersive, meaning that their optical characteristics change as a function of wavelength. For example, a power splitting ratio of an MZI can change over wavelength. However, by designing the optoelectronic matrix multiplication unit 3520 to have a sufficiently large operating wavelength window, and by limiting the wavelengths to be within that operating wavelength window, the output electronic vector output by the optoelectronic matrix multiplication unit 3520 corresponding to each wavelength can be a sufficiently accurate result of the matrix multiplication implemented by the optoelectronic matrix multiplication unit 3520. The operating wavelength window can be, for example, 1 nm, 2 nm, 3 nm, 4 nm, 5 nm, 10 nm, or 20 nm.
The modulator array 144 of the WDM ANN computation system 3500 includes banks of optical modulators configured to generate a plurality of optical input vectors, each of the banks corresponding to one of the multiple wavelengths and generating respective optical input vector having respective wavelength. For example, for a system with an optical input vector of length 32 and 3 wavelengths (e.g., λ1, λ2, and λ3), the modulator array 144 can have 3 banks of 32 modulators each. Further, the modulator array 144 also includes an optical multiplexer configured to combine the plurality of optical input vectors into a combined optical input vector including the plurality of wavelengths. For example, the optical multiplexer can combine the outputs of the three banks of modulators at three different wavelengths into a single propagation channel, such as a waveguide, for each element of the optical input vector. As such, returning to the example above, the combined optical input vector would have 32 optical signals, each signal containing 3 wavelengths.
The optoelectronic processing components of the WDM ANN computation system 3500 are further configured to demultiplex the multiple wavelengths and to generate a plurality of demultiplexed output electric signals. Referring to
Copies of the optical input vector element v2 at the wavelengths λ1, λ2, and λ3 are provided to the multiplication module 3530_12, 3530_22, . . . , and 3530_m2. The multiplication module 3530_12 outputs three electrical signals representing M12·v2 that correspond to the input vector element v2 at the wavelengths λ1, λ2, and λ3. The multiplication module 3530_22 outputs three electrical signals representing M22·v2 that correspond to the input vector element v2 at the wavelengths λ1, λ2, and λ3. The multiplication module 3530_m2 outputs three electrical signals representing Mm2·v2 that correspond to the input vector element v2 at the wavelengths λ1, λ2, and λ3.
Copies of the optical input vector element vn including the wavelengths λ1, λ2, and λ3 are provided to the multiplication module 3530_1n, 3530_2n, . . . , and 3530_mn. The multiplication module 3530_1n outputs three electrical signals representing M1n·vn that correspond to the input vector element vn at the wavelengths λ1, λ2, and λ3. The multiplication module 3530_2n outputs three electrical signals representing M2n·vn that correspond to the input vector element vn at the wavelengths λ1, λ2, and λ3. The multiplication module 3530_mn outputs three electrical signals representing Mmn·vn that correspond to the input vector element vn at the wavelengths λ1, λ2, and λ3, and so forth.
For example, each of the multiplication module 3530 can include a demultiplexer configured to demultiplex the three wavelengths contained in each of the 32 signals of the multi-wavelength optical vector, and route the 3 single-wavelength optical output vectors to three banks of photodetectors (e.g., photodetectors 2012, 2016 (
Three banks of summation modules 1808 receive outputs from the multiplication modules 3530 and generate sums y that correspond to the input vector at the various wavelengths, For example, three summation modules 1808_1 receive the outputs of the multiplication modules 3530_11, 3530_12, . . . , 3530_1n and generate sums y1(λ1), y1(λ2), y1(λ3) that correspond to the output vector element y1 at the wavelengths λ1, λ2, and λ3, respectively, in which at each wavelength the sum y1 is equal to M11v1+M12v2+ . . . +M1nvn. Three summation modules 1808_2 receive the outputs of the multiplication modules 3530_21, 3530_22, . . . , 3530_2n, and generates sums y2(λ1), y2(λ2), y2(λ3) that correspond to the output vector element y2 at the wavelengths λ1, λ2, and λ3, respectively, in which at each wavelength the sum y2 is equal to M21v1+M22v2+ . . . +M2nvn. Three summation modules 1808_m receive the outputs of the multiplication modules 3530_m1, 3530_m2, . . . , 3530_mn, and generates sums ym(λ1), ym(λ2), ym(λ3) that correspond to the output vector element ym at the wavelengths λ1, λ2, and λ3, respectively, in which at each wavelength the sum ym is equal to Mm1+Mm2 v2+ . . . +Mmnvn.
Referring back to
The controller 110 can implement a method analogous to the method 200 (
In some cases, the ANN can be specifically designed, and the digital input vectors can be specifically formed such that the multi-wavelength products of the multiplication module 3530 can be added without demultiplexing. In such cases, the multiplication module 3530 can be a wavelength-insensitive multiplication module that does not demultiplex the multiple wavelengths of the multi-wavelength products. As such, each of the photodetectors of the multiplication module 3530 effectively sums the multiple wavelengths of an optical signal into a single photocurrent, and each of the voltages output by the multiplication module 3530 corresponds to a sum of the product of a vector element and a matrix element for the multiple wavelengths. The summation module 1808 (only one bank is needed) outputs an element-by-element sum of the matrix multiplication results of the multiple digital input vectors.
and the matrix is
In this example, the input vector has multiple wavelengths λ1, λ2, and λ3, and each of the elements of the input vector is encoded on a different optical signal. Two different copying modules 1902 perform an optical copying operation to split the computation over different paths (e.g., an “upper” path and a “lower” path). There are four multiplication modules 1904 that each multiply by a different matrix element using optical amplitude modulation. The output of each multiplication module 1904 is provided to a demultiplexer and a bank of optical detection modules 3502 that convert a wavelength division multiplexed optical signal to electrical signals in the form of electrical currents associated with the wavelengths λ1, λ2, and λ3. Both upper paths of the different input vector elements are combined using a bank of summation modules 3504 associated with the wavelengths λ1, λ2, and λ3, and both lower paths of the different input vector elements are combined using a bank of summation modules 3504 associated with the wavelengths λ1, λ2, and λ3, in which the summation modules 3504 perform summation in the electrical domain. Thus, each of the elements of the output vector for each wavelength is encoded on a different electrical signal. As shown in
M11v1+M12v2
M21v1+M22v2
The system configuration 3500 can be implemented using any of a variety of optoelectronic technologies. In some implementations, there is a common substrate (e.g., a semiconductor such as silicon), which can be able to support both integrated optics components and electronic components. The optical paths can be implemented in waveguide structures that have a material with a higher optical index surrounded by a material with a lower optical index defining a waveguide for propagating an optical wave that carries an optical signal. The electrical paths can be implemented by a conducting material for propagating an electrical current that carries an electrical signal. (In
A variety of alternative system configurations or signal processing techniques can be used with various implementations of the different systems, subsystems, and modules described herein.
In some embodiments, it can be useful for some or all of the VMM subsystems to be replaceable with alternative subsystems, including subsystems that use different implementations of the various copying modules, multiplication modules, and/or summation modules. For example, a VMM subsystem can include the optical copying modules described herein and the electrical summation modules described herein, but the multiplication modules can be replaced with a subsystem that performs the multiplication operations in the electrical domain instead of the optoelectronic domain. In such examples, the array of optical amplitude modulators can be replaced by an array of detectors to convert optical signals to electrical signals, followed by an electronic subsystem (e.g., an ASIC, processor, or SoC). Optionally, if optical signal routing is to be used to the summation modules that are configured to detect optical signals, the electronic subsystem can include electrical to optical conversion, for example, using an array of electrically-modulated optical sources.
In some embodiments, it can be useful to be able to use a single wavelength for some or all of the optical signals being used for some or all of the VMM computations. Alternatively, in some embodiments, to help reduce the number of required optical input ports, an input port can receive a multiplexed optical signal that has different values encoded on different optical waves at different wavelengths. Those optical waves can then be separated at an appropriate location in the system, depending on whether any of the copying modules, multiplication modules, and/or summation modules are configured to operate on multiple wavelengths. But, even in the multi-wavelength embodiments, it can be useful to use the same wavelength for different subsets of optical signals, for example, used in the same VMM subsystem.
In some embodiments, an accumulator can be used to enable a time domain encoding of the optical and electrical signals received by the various modules, alleviating the need for the electronic circuitry to operate effectively over a large number of different power levels. For example, a signal that is encoded using binary (on-off) amplitude modulation with a particular duty cycle over N time slots per symbol, can be converted into a signal that has N amplitude levels per symbol after that signal is passed through the accumulator (an analog electronic accumulator that integrates the current or voltage of an electrical signal). So, if the optical devices (e.g., the phase modulators in the optical amplitude modulators) are capable of operating at a symbol bandwidth B, they can be operated instead at a symbol bandwidth B/100, where each symbol value uses N=100 time slots. An integrated amplitude of 50% has a 50% duty cycle (e.g., the first 50 time slots at the non-zero “on” level, followed by 50 time slots at the zero, or near zero, “off” level), whereas an integrated amplitude of 10% has a 10% duty cycle (e.g., the first 10 time slots at the non-zero “on” level, followed by 90 time slots at zero “off” level). In the examples described herein, such an accumulator can be positioned on the path of each electrical signal at any location within the VMM subsystem that is consistent for each electrical signal, such as for example, before the summation modules for all electrical signals in that VMM subsystem or after the summation modules for all electrical signals in that VMM subsystem. The VMM subsystem can also be configured such that there are no significant relative time shifts between different electrical signals preserving alignment of the different symbols.
Referring to
For example, the homodyne detector 4000 can be used in the systems shown in
In the example of
A way to reduce the number of optical fibers required to carry optical signals to an optoelectronics chip is to use wavelength division multiplexing. Multiple optical signals having different wavelengths can be multiplexed and transmitted using a single optical fiber. For example, referring to
Inside the optoelectronics chip 4114, the wavelength division multiplexed signal is demultiplexed by a demultiplexer 4302 to separate the optical signals 4120 and 4122. In this example, the optical signal 4120 is copied by a copying module 4304 to produce copies of optical signals that are sent to the matrix multiplication modules 4116a and 4118a. The optical signal 4122 is copied by a copying module 4306 to produce copies of optical signals that are sent to the matrix multiplication modules 4116b and 4118b. The outputs of the matrix multiplication units 4116a and 4116b are combined using an optical coupler 4120a, and the combined signal is detected by a photodetector 4122a.
A third light signal 4124 having a wavelength λ1 is modulated by a third modulator 4128 to produce a third modulated optical signal 4132 representing a third input vector element V3. A fourth light signal 4126 having a wavelength λ2 is modulated by a fourth modulator 4130 to produce a fourth modulated optical signal 4134 representing a fourth input vector element V4. The third and fourth modulated optical signals are combined by a multiplexer 4136 to produce a wavelength division multiplexed signal that is transmitted via an optical fiber 4138 to the optoelectronics chip 4114.
Inside the optoelectronics chip 4114, the wavelength division multiplexed signal provided by the optical fiber 4138 is demultiplexed by a demultiplexer 4140 to separate the optical signals 4132 and 4134. In this example, the optical signal 4132 is copied by a copying module 4142 to produce copies of optical signals that are sent to the matrix multiplication modules 4116c and 4118c. The optical signal 4134 is copied by a copying module 4144 to produce copies of optical signals that are sent to the matrix multiplication modules 4116d and 4118d. The outputs of the matrix multiplication units 4116c and 4116d are combined using an optical coupler 4120b, and the combined signal is detected by a photodetector 4122b. The outputs of the matrix multiplication units 4118a and 4118b are combined using an optical coupler, and the combined signal is detected by a photodetector. The outputs of the matrix multiplication units 4118c and 4118d are combined using an optical coupler, and the combined signal is detected by a photodetector.
In some examples, a multiplexer can multiplex optical signals having three or more (e.g., 10, or 100) wavelengths to produce a wavelength division multiplexed signal that is transported by a single optical fiber, and a demultiplexer inside the optoelectronics chip can demultiplex the wavelength division multiplexed signal to separate the signals having different wavelengths. This allows more optical signals be transmitted to the optoelectronics chip in parallel through the optical fibers, increasing the data processing throughput of the optoelectronics chip.
In some examples, the laser unit 142 of
The optical interference unit 154 can include a plurality of interconnected Mach-Zehnder interferometers (MZIs).
In the examples shown in
During operation of the system, light propagates from the input waveguide 4704 to the 1×2 MMI splitter 4708, which generates two light portions. Each light portion is influenced by the electro-optic phase shifters 4714, then the two light portions combine at the 1×2 MMI 4710 to interfere so that optical intensity at the output waveguide 4706 will have a relationship with the relative phase difference between the two arms 4712a and 4712b. For example, the modulation phase shifters 4714 can be driven by a plurality of on-off electrical control signals that are generated by flip chip electrical circuits. By implementing different lengths of the phase shifters 4714, output light can encode multiple levels of electrical signals through several binary on-off keying digital inputs.
The upper arm waveguide 4712a and the lower arm waveguide 4712b can differ in waveguide width due to the fabrication process, and the waveguide width variation induced phase difference will accumulate along the propagation direction between the two arms so that the initial phase imbalance between the two arms 4712a, 4712b can make the MZI initial operating point far away from constructive inference. In addition, the plurality of electro-optic phase shifters 4714 can experience mask misalignment in fabrication that will also induce initial phase difference between the two arms 4712a, 4712b. The two calibration phase shifters 4702a, 4702b are used to cancel the fabrication induced phase imbalance.
For example, the calibration phase shifters 4702 can be implemented by thermo-optic phase effects (operated by heating the waveguide to change the refractive index) or electro-optic effects (operated by applying an electric field to change carrier distribution to influence the refractive index). For example, the calibration phase shifters 4702 are independently controlled by two low speed DACs through flip chip or wire bonding to the photonic chip that includes the MZI 4700.
For example, a tap waveguide is provided to guide a portion of the light signal at the output waveguide 4706 to a monitor photodetector that can provide a feedback signal indicating the intensity of the light propagating in the output waveguide 4706. To calibrate the MZI 4700, the phase shifters 4714 are driven with zero signals so that any imbalance between the phase of the light portions in the upper and lower arms is due to the differences between the waveguides 4712a, 4712b in the upper and lower arms. Based on the feedback provided by the monitor photodetector, the control signals applied to the calibration phase shifters 4702a, 4702b are adjusted so that the light portions from the upper and lower arms that reach the output 1×2 MMI splitter 4710 have the same phase and constructively interfere.
The phase difference between the upper and lower arms can be affected by ambient temperature, which can fluctuate over time, so the measurements by the monitor photodetector are performed periodically, and the control signals applied to the calibration phase shifters 4702 are updated periodically, e.g., once every fraction of a second, once very second, once every 10 seconds, or once every minute. The calibration phase shifters 4702 are configured to be able to compensate for phase imbalances in the two arms ranging from 0 to a radians. Because the control signals to the calibration phase shifters 4702 can be updated at a low frequency (e.g., 10 Hz or less), a high-precision multi-level DAC can be used to generate the control signals for the calibration phase shifters 4702.
Referring to
Referring to
In general, the electro-optic phase shifters for converting the digital electric signals to analog optical signals can be placed on a single arm of the MZI, or on both arms of the MZI. The MZI can include a single calibration phase shifter placed on a single arm, or include two calibration phase shifters placed on both arms, for compensating the phase imbalance between the two arms. In general, having phase shifters placed in both arms of the MZI allows each phase shifter to be shorter since each phase shifter only has to impart a smaller amount of optical phase shift, as compared to placing the phase shifters in only one arm of the MZI.
For example, the modulator that uses the segmented design in
In this example, the laser unit 142 outputs four light signals on four optical waveguides 4704a, 4704b, 4704c, and 4704d. The light signal on the waveguide 4704a is modulated by the optical modulator 5002 according to the first 4-bit value of the digital input vector. The light signal on the waveguide 4704b is modulated by the optical modulator 5004 according to the second 4-bit value of the digital input vector. The light signal on the waveguide 4704c is modulated by the optical modulator 5006 according to the third 4-bit value of the digital input vector. The light signal on the waveguide 4704d is modulated by the optical modulator 5008 according to the fourth 4-bit value of the digital input vector.
The optical modulator 5002 includes two 1×2 port multi-mode interference couplers (MMI 1×2) 4708a and 4710a, two balanced arms 5012a and 5012b, and eight sub-modulators: 5022a, 5022b, 5022c, 5022d in the upper arm and 5022e, 5022f, 5022g, 5022h in the lower arm. For example, each sub-modulator can include a phase shifter. The sub-modulators 5022d and 5022h receive the LSB (B11) of the first value in the input vector. The sub-modulators 5022c and 5022g receive the 2nd-bit (B12) of the first value in the input vector. The sub-modulators 5022b and 5022f receive the 3rd-bit (B13) of the first value in the input vector. The sub-modulators 5022a and 5022e receive the MSB (B14) of the first value in the input vector. The optical modulator 5002 also includes calibration phase shifters 5032a and 5032b that are driven by multi-bit DACs.
The sub-modulator 5022d includes a first waveguide segment, the sub-modulator 5022c includes a second waveguide segment, the sub-modulator 5022b includes a third waveguide segment, and the sub-modulator 5022a includes a fourth waveguide segment. The second waveguide segment is approximately twice as long as the first waveguide segment, the third waveguide segment is approximately four times as long as the first waveguide segment, and the fourth waveguide segment is approximately eight times as long as the first waveguide segment.
Similarly, the sub-modulator 5022h includes a fifth waveguide segment, the sub-modulator 5022g includes a sixth waveguide segment, the sub-modulator 5022f includes a seventh waveguide segment, and the sub-modulator 5022e includes an eighth waveguide segment. The sixth waveguide segment is approximately twice as long as the fifth waveguide segment, the seventh waveguide segment is approximately four times as long as the fifth waveguide segment, and the eighth waveguide segment is approximately eight times as long as the fifth waveguide segment.
The optical modulator 5004 includes two 1×2 port multi-mode interference couplers (MMI 1×2) 4708b and 4710b, two balanced arms 5014a and 5014b, and eight sub-modulators: 5024a, 5024b, 5024c, 5024d in the upper arm and 5024e, 5024f, 5024g, 5024h in the lower arm. For example, each sub-modulator can include a phase shifter. The sub-modulators 5024d and 5024h receive the LSB (B21) of the second value in the input vector. The sub-modulators 5024c and 5024g receive the 2nd-bit (B22) of the second value in the input vector. The sub-modulators 5024b and 5024f receive the 3rd-bit (B23) of the second value in the input vector. The sub-modulators 5024a and 5024e receive the MSB (B24) of the second value in the input vector. The optical modulator 5004 also includes calibration phase shifters 5034a and 5034b that are driven by multi-bit DACs.
The optical modulators 5006 and 5008 are configured in a manner similar to the optical modulators 5002 and 5004.
The optical modulator 5002 modulates the optical signal in the input waveguide 4708a according to the four bits of the first value in the digital input vector, and generates an analog optical signal 5042 that represents the first value of the digital input vector. The optical modulator 5002 converts the first 4-bit value of the digital input vector to a first analog optical signal. The optical modulator 5004 modulates the optical signal in the input waveguide 4708b according to the four bits of the second value in the digital input vector, and generates an analog optical signal 5044 that represents the second value of the digital input vector. The optical modulator 5004 converts the second 4-bit value of the digital input vector to a second analog optical signal.
The optical modulator 5006 modulates the optical signal in the input waveguide 4708c according to the four bits of the third value in the digital input vector, and generates an analog optical signal 5046 that represents the third value of the digital input vector. The optical modulator 5006 converts the third 4-bit value of the digital input vector to a third analog optical signal. The optical modulator 5008 modulates the optical signal in the input waveguide 4708d according to the four bits of the fourth value in the digital input vector, and generates an analog optical signal 5048 that represents the fourth value of the digital input vector. The optical modulator 5008 converts the fourth 4-bit value of the digital input vector to a fourth analog optical signal. The analog optical signals 5042, 5044, 5046, 5048 together form an analog optical input vector that can be provided to, e.g., the optoelectronic matrix multiplication unit 150 of
The ANN computation systems described above can be used in, e.g., personal computers, portable computers, mobile phones, tablet computers. The ANN computation systems described above can be implemented as a peripheral card that can be electrically coupled to a computer motherboard. The peripheral card can be, e.g., a PCIe card that complies with PCIe 3.0, PCIe 4.0, or later versions of the PCIe standard. The ANN computation systems can be used in, e.g., server computers, autonomous driving vehicles, autonomous trucks, autonomous trains, autonomous drones, autonomous ships, autonomous airplanes, autonomous space ships, satellites, data centers, and supercomputers.
The digital controller (e.g., for controlling the components shown in
A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program does not necessarily correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.
The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
Some background information for the various systems described in this specification, and additional details of some of the modules the various systems described in this specification, are provided in U.S. patent application Ser. No. 16/431,167, filed on Jun. 4, 2019, U.S. patent application Ser. No. 17/204,320, filed on Mar. 17, 2021, U.S. provisional application 63/048,439, filed on Jul. 6, 2020, U.S. provisional application 63/061,995, filed on Aug. 6, 2020, and U.S. provisional application 63/123,338, filed on Dec. 9, 2020. The above patent applications and provisional applications are incorporated herein by reference.
While the disclosure has been described in connection with certain embodiments, it is to be understood that the disclosure is not to be limited to the disclosed embodiments but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the scope of the appended claims, which scope is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures as is permitted under the law.
This application claims priority to U.S. Provisional Application 63/017,211, filed on Apr. 29, 2020, the entire disclosure of which is incorporated by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 3944820 | Stotts | Mar 1976 | A |
| 4567569 | Caulfield et al. | Jan 1986 | A |
| 4588255 | Tur et al. | May 1986 | A |
| 4603398 | Bocker et al. | Jul 1986 | A |
| 4633428 | Byron | Dec 1986 | A |
| 4809204 | Dagenais et al. | Feb 1989 | A |
| 5004309 | Caulfield et al. | Apr 1991 | A |
| 5008833 | Agranat et al. | Apr 1991 | A |
| 5077619 | Toms | Dec 1991 | A |
| 5095459 | Ohta et al. | Mar 1992 | A |
| 5220643 | Collings | Jun 1993 | A |
| 5428711 | Akiyama et al. | Jun 1995 | A |
| 5671302 | Skeie | Sep 1997 | A |
| 5699449 | Javidi | Dec 1997 | A |
| 5784309 | Budil | Jul 1998 | A |
| 6005998 | Lee | Dec 1999 | A |
| 6178020 | Schultz et al. | Jan 2001 | B1 |
| 6307854 | Webb | Oct 2001 | B1 |
| 6493483 | Gomes et al. | Dec 2002 | B2 |
| 7173272 | Ralph | Feb 2007 | B2 |
| 7251386 | Dickinson et al. | Jul 2007 | B1 |
| 7259031 | Dickinson et al. | Aug 2007 | B1 |
| 7536431 | Goren et al. | May 2009 | B2 |
| 7660533 | Meyers et al. | Feb 2010 | B1 |
| 7667995 | Leuenberger et al. | Feb 2010 | B1 |
| 7876248 | Berkley et al. | Jan 2011 | B2 |
| 7985965 | Barker et al. | Jul 2011 | B2 |
| 8018244 | Berkley | Sep 2011 | B2 |
| 8023828 | Beausoleil et al. | Sep 2011 | B2 |
| 8027587 | Watts | Sep 2011 | B1 |
| 8035540 | Berkley et al. | Oct 2011 | B2 |
| 8129670 | Laycock et al. | Mar 2012 | B2 |
| 8190553 | Routt | May 2012 | B2 |
| 8213751 | Ho et al. | Jul 2012 | B1 |
| 8223414 | Goto et al. | Jul 2012 | B2 |
| 8386899 | Goto et al. | Feb 2013 | B2 |
| 8560282 | Macready et al. | Oct 2013 | B2 |
| 8604944 | Berkley et al. | Dec 2013 | B2 |
| 8620855 | Bonderson | Dec 2013 | B2 |
| 8837544 | Santori et al. | Sep 2014 | B2 |
| 8953950 | Nazarathy et al. | Feb 2015 | B2 |
| 9159861 | Zhang et al. | Oct 2015 | B2 |
| 9250391 | McLaughlin et al. | Feb 2016 | B2 |
| 9354039 | Mower et al. | May 2016 | B2 |
| 9391708 | Fincato et al. | Jul 2016 | B2 |
| 9594394 | New | Mar 2017 | B2 |
| 9791258 | Mower et al. | Oct 2017 | B2 |
| 9858531 | Monroe et al. | Jan 2018 | B1 |
| 9900096 | Hajimiri et al. | Feb 2018 | B2 |
| 10009135 | Tait et al. | Jun 2018 | B2 |
| 10268232 | Carolan et al. | Apr 2019 | B2 |
| 10359272 | Mower et al. | Jul 2019 | B2 |
| 10608663 | Gould et al. | Mar 2020 | B2 |
| 10634851 | Steinbrecher et al. | Apr 2020 | B2 |
| 10908634 | Chang et al. | Feb 2021 | B1 |
| 11474409 | Sobu | Oct 2022 | B2 |
| 20030086138 | Pittman et al. | May 2003 | A1 |
| 20030235363 | Pfeiffer | Dec 2003 | A1 |
| 20040223768 | Shastri et al. | Nov 2004 | A1 |
| 20040243657 | Goren et al. | Dec 2004 | A1 |
| 20070180586 | Amin | Aug 2007 | A1 |
| 20080031566 | Matsubara et al. | Feb 2008 | A1 |
| 20080037936 | Lee | Feb 2008 | A1 |
| 20080212186 | Zoller et al. | Sep 2008 | A1 |
| 20080273835 | Popovic | Nov 2008 | A1 |
| 20090024816 | Thibadeau et al. | Jan 2009 | A1 |
| 20090028554 | Anderson | Jan 2009 | A1 |
| 20090128228 | Yeh | May 2009 | A1 |
| 20110064419 | Chen | Mar 2011 | A1 |
| 20120045162 | Li | Feb 2012 | A1 |
| 20120171619 | Heyderman et al. | Jul 2012 | A1 |
| 20130011093 | Goh et al. | Jan 2013 | A1 |
| 20130046716 | Chan et al. | Feb 2013 | A1 |
| 20140241657 | Manouvrier | Aug 2014 | A1 |
| 20140299743 | Miller | Oct 2014 | A1 |
| 20150009548 | Bienstman et al. | Jan 2015 | A1 |
| 20150354938 | Mower et al. | Dec 2015 | A1 |
| 20150382089 | Mazed | Dec 2015 | A1 |
| 20160103281 | Matsumoto | Apr 2016 | A1 |
| 20160118106 | Yoshimura et al. | Apr 2016 | A1 |
| 20160139485 | Winzer | May 2016 | A1 |
| 20160162798 | Marandi et al. | Jun 2016 | A1 |
| 20170285373 | Zhang et al. | Oct 2017 | A1 |
| 20170351293 | Carolan | Dec 2017 | A1 |
| 20180024299 | Leijtens et al. | Jan 2018 | A1 |
| 20180059445 | Miyazaki | Mar 2018 | A1 |
| 20180107237 | Andregg et al. | Apr 2018 | A1 |
| 20180267937 | Pelc et al. | Sep 2018 | A1 |
| 20180274900 | Mower et al. | Sep 2018 | A1 |
| 20180335574 | Steinbrecher et al. | Nov 2018 | A1 |
| 20180343064 | Kawaai | Nov 2018 | A1 |
| 20190019100 | Roques-Carmes et al. | Jan 2019 | A1 |
| 20190049680 | Sharma et al. | Feb 2019 | A1 |
| 20190094648 | Williams et al. | Mar 2019 | A1 |
| 20190227351 | Behroozpour | Jul 2019 | A1 |
| 20190266508 | Bunyk et al. | Aug 2019 | A1 |
| 20190318234 | Abel et al. | Oct 2019 | A1 |
| 20190370652 | Shen et al. | Dec 2019 | A1 |
| 20190371226 | Iwaki | Dec 2019 | A1 |
| 20200018193 | Neiser | Jan 2020 | A1 |
| 20200110992 | Hosseinzadeh et al. | Apr 2020 | A1 |
| 20200228077 | Harris et al. | Jul 2020 | A1 |
| 20210173238 | Hosseinzadeh et al. | Jun 2021 | A1 |
| 20210201126 | Meng et al. | Jul 2021 | A1 |
| Number | Date | Country |
|---|---|---|
| 1713027 | Dec 2005 | CN |
| 101630178 | Jan 2010 | CN |
| 102023672 | Apr 2011 | CN |
| 102164017 | Aug 2011 | CN |
| 102866876 | Jan 2013 | CN |
| 103473213 | Dec 2013 | CN |
| 105046325 | Nov 2015 | CN |
| 109477938 | Mar 2019 | CN |
| 102004014658 | Jul 2010 | DE |
| 0 594 088 | Dec 1988 | EP |
| 0 330 710 | Sep 1989 | EP |
| 10-2017- 0117557 | Oct 2017 | KR |
| 256884 | Sep 1995 | TW |
| 384403 | Mar 2000 | TW |
| 200912905 | Mar 2009 | TW |
| I522935 | Feb 2016 | TW |
| 201837894 | Oct 2018 | TW |
| 202001694 | Jan 2020 | TW |
| WO 2005029404 | Mar 2005 | WO |
| WO 2006023067 | Mar 2006 | WO |
| WO 2008069490 | Jun 2008 | WO |
| WO 2018098230 | May 2018 | WO |
| WO 2019236591 | Dec 2019 | WO |
| WO 2020191217 | Sep 2020 | WO |
| Entry |
|---|
| Abdeldayem et al., “Optical Computing: Need and Challenge”, Communications of the ACM, vol. 50, No. 9. pp. 60-62 (Sep. 2007). |
| The International Search Report and Written Opinion of the International Searching Authority for International Application No. PCT/US2019/035403 dated Sep. 18, 2019. |
| The International Search Report and Written Opinion of the International Searching Authority for International Application No. PCT/US2019/064554 dated Jun. 23, 2020. |
| The International Search Report and Written Opinion of the International Searching Authority for International Application No. PCT/US2020/023674 dated Jul. 8, 2020. |
| The Taiwan Office Action for Taiwan Application No. TW 108119388 dated May 13, 2020 (with English Translation). |
| Debnath, Kapil et al., “All-silicon carrier accumulation modulator based on a lateral metal-oxide-semiconductor capacitor”, Photonics Research, vol. 6, No. 5, pp. 373-379 (May 2018). |
| Green, William M.J. et al., “Ultra-compact, lowRF power, 10 GB/s silicon Mach-Zehnder modulator”, Optics Express, vol. 15, No. 25, pp. 17106-17113 (Dec. 10, 2007). |
| Guha et al., “Architectural Issues in Designing Symbolic Processors in Optics”, ISCA '87 Proceedings of the 14th annual international symposium on Computer Architecture, pp. 145-151 (1987). |
| Hamerly et al., “Large-Scale Optical Neural Networks based on Photoelectric Multiplication”, arXiv: 1812.07614v1 (cs.ET] (Nov. 12, 2018). |
| Harris et al., “Linear programmable nanophotonic processors”, Optica, vol. 5, No. 12, pp. 1623-1631 (Dec. 2018). |
| Huang, K., “A Digital Optical Cellular Image Processor (DOCIP): Theory, Architecture and Implementation”, A Dissertation Presented to the Faculty of the Graduate School University of Southern California, (Sep. 1988). |
| Hughes et al., “Training of photonic neural networks through in situ backpropagation”, Optica, vol. 5, Issue 7, pp. 864-871 (2018). |
| Ishihara et al., “An Integrated Nanophotonic Parallel Adder”, ACM Journal on Emerging Technologies in Computing Systems, vol. 14, No. 2, Article 26, pp. 26:1-26:20 (Jul. 2018). |
| Jing et al., “Tunable Efficient Unitary Neural Networks (EUNN) and their application toRNNs”, https://arxiv.org/abs/1612.05231v3, (Dec. 15, 2016). |
| Lin et al., “All-Optical Machine Learning Using Diffractive Deep Neural Networks”, Science, vol. 361, Issue 6406, pp. 1004-1008, DOI: 10,1126/science.aat8084 (Sep. 7, 2018). |
| Patel, David et al., “Silicon Photonic Segmented Modulator-Based Electro-Optic DAC for 100 GB/s PAM-4 Generation”, IEEE Photonics Technology Letters, vol. 27, No. 23, pp. 2433-2435 (Dec. 1, 2015). |
| Peurifoy et al., “Nanophotonic particle simulation and inverse design using artificial neural networks”, Science Advances, vol. 4, eaar4206, pp. 1-7 (2018). |
| Saade et al., “Random Projections through multiple optical scattering: Approximating kernels at the speed of light”, Proceedings of the 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6215-6219 (Oct. 2015). |
| Sahni et al., “Models and Algorithms for Optical and Optoelectronic Parallel Computers”, International Journal of Foundations of Computer Science, vol. 12, No. 3, (Nov. 2011). |
| Shen et al., “Deep learning with coherent nanophotonic circuits”, Nature Photonics, DOI: 10.103 8/NPHOTON.2017.93 (Jun. 12, 2017). |
| Tahersima, M. et al., “Deep Neural Network Inverse Design of Integrated Photonic Power Splitters”, Scientific Reports, Article No. 9:1368, (Feb. 4, 2019). |
| Tamir et al., “High-speed and low-power electro-optical DSP coprocessor”, J. Opt. Soc. Am. A, vol. 26, No. 8, pp. A11-A20 (Aug. 2009). |
| Wu, Xiaotie et al., “A 20Gb/s NRZ/PAM-4 1 V Transmitter in 40nm CMOS Driving a Si-Photonic Modulator In 0.13μm CMOS”, 2013 IEEE International Solid-State Circuits Conference, pp. 128-129 (Feb. 19, 2013). |
| Xu, K. et al., “All passive photonic power divider with arbitrary split ratio”, arXiv, 1609,03 823v1 (Sep. 13, 2016). |
| Yang et al., “On-chip CMOS-compatible optical signal processor”, Optics Express, vol. 20, No. 12, pp. 13560-13565 (Jun. 4, 2012). |
| Younger et al., “Computing by Means of Physical-Based Optical Neural Networks”, EPTCS 26, pp. 159-167 (2010). |
| Aaronson, S. et al., “Computational complexity of linear optics”, in Proceedings of the 43rd Annual ACM Symposium on Theory of Computing (ACM. New York,.Ny, USA, 2011), STOC '11, pp. 333-342, ISBN 978-1-4503-0691-1. |
| Abu-Mostafa et al., “Optical Neural Computers,” Scientific American, vol. 256, Issue 3 (1987), pp. 88-95. |
| Albert et al., “Statistical mechanics of complex networks”, Reviews of Modern Physics, vol. 74, pp. 47-97, Jan. 2002. |
| Almeida, V. R., et al., “All-optical control of light on a silicon chip”, Nature, vol. 431, (Aug. 6, 2004), pp. 1081-1084. |
| Amir, A. et al., “Classical diffusion of a quantum particle in a noisy environment”, Physical Review, E 79, 050105 (Feb. 5, 2009), 5 pages. |
| Amit et al., “Spin-glass models of neural networks”, Physical Review A, vol. 32, No. 2, pp. 1007-1018, Aug. 1985. |
| Anitha et al., “Comparative Study of High performance Braun's Multiplier Using FPGAs”, IOSR Journal of Electronics and Communication Engineering (IOSRJECE), vol. 1, Issue 4, pp. 33-37 (May-Jun. 2012). |
| Appellant et al., “Information processing using a single dynamical node as complex system”, Nature Communications, 6 pages (2011). |
| Arjovsky et al., “Unitary Evolution Recurrent Neural Networks”, arXiv:1511.06464, 9 pages (2015). |
| Aspuru-Guzik A. et al., “Simulated Quantum Computation of Molecular Energies”, Science, vol. 309, Issue 5741, pp. 1704-1707 (2005). |
| Aspuru-Guzik, A. et al., “Photonic quantum simulators”, Nature Physics, vol. 8, No. 4, pp. 285-291 (2012). |
| Atabaki et al., “Integrating photonics with silicon nanoelectronics for the next generation of systems on a chip”, Nature, vol. 556, pp. 349-358 (2018). |
| Baehr-Jones et al., “A 25 GB/s Silicon Photonics Platform”, arXiv reprints. URL http://adsabs.harvard.edu/ abs/2012arXivl203.0767B, 1203.0767, 11 pages (2012). |
| Bao et al., “Atomic-Layer Graphene as a Saturable Absorber for Ultrafast Pulsed Lasers”, Advanced Functional Materials, vol. 19, pp. 3077-3083 (2009). |
| Bao et al., “Monolayer graphene as a saturable absorber in a mode-locked laser”, Nano Res., vol. 4, No. 3, pp. 297-307 (2011). |
| Barahona, “On the computational complexity of Ising spin glass models”, Journal of Physics A: Mathematical and General, vol. 15, pp. 3241-3253, (Oct. 1982). |
| Bertsimas et al., “Robust optimization with simulated annealing”, Journal of Global Optimization, vol. 48, pp. 323-334 (2010). |
| Bewick, “Fast multiplication: algorithms and implementation”. Ph.D. Thesis, Stanford University (1994), 170 pages. |
| Bonneau et al., “Quantum interference and manipulation of entanglement in silicon wire waveguide quantum circuits”, New Journal of Physics, vol. 14 (2012): 045003, 13 pages. |
| Brilliantov, “Effective magnetic Hamiltonian and Ginzburg criterion for fluids”, Physical Review E, vol. 58, No. 2, pp. 2628-2631, (Aug. 1998). |
| Bromberg, Y. et al., “Bloch oscillations of path-entangled photons”, Phys. Rev. Lett., vol. 105, (May 18, 2011), 5 pages. |
| Bromberg, Y. et al., “Quantum and Classical Correlations in Waveguide Lattices”, Phys. Rev. Lett., vol. 102, (Jun. 26, 2009), pp. 253904-1-253904-4. |
| Broome, M. A. et al., “Photonic Boson Sampling in a Tunable Circuit”, Science, vol. 339, pp. 794-798, (Jan. 12, 2013). |
| Bruck et al., “On the power of neural networks for solving hard problems”, Journal of Complexity, vol. 6, pp. 129-135, (Jun. 1990). |
| Canziani et al., “Evaluation of neural network architectures for embedded systems”, 2017 IEEE International Symposium on Circuits and Systems (ISCAS), (May 28-31, 2017). |
| Cardenas et al., “Low loss etchless silicon photonic waveguides”,Optics Express, vol. 17, No. 6, pp. 4752-4757 (2009). |
| Carolan et al., “Universal linear optics”, Science, vol. 349, Issue 6249, pp. 711-716, (Aug. 14, 2015). |
| Caves, “Quantum-mechanical noise in an interferometer”, Physical Review D, vol. 23, No. 8, pp. 1693-1708 (Apr. 15, 1981). |
| Centeno et al., “Optical bistability in finite-size nonlinear bidimensional photonic crystals doped by a microcavity”, Physical Review B, vol. 62, No. 12, pp. R7683-R7686 (Sep. 15, 2000). |
| Chan, “Optical flow switching networks”, Proceedings of the IEEE, vol. 100, No. 5, pp. 1079-1091, (May 2012). |
| Chen et al., “DianNao: A small-footprint high-throughput accelerator for ubiquitous machine-learning”, ASPLOS '14: Proceedings of the 19th international conference on Architectural support for programming languages and operating systems, pp. 269-284 (Febr. |
| Chen, J. et al., “Efficient photon pair sources based on silicon-on-insulator microresonators”, SPIE, vol. 7815, pp. 78150H-1-78150H-9, (2010). |
| Chen, J. et al., “Frequency-bin entangled comb of photon pairs from a Silicon-on-Insulator micro-resonator”, Optics Express, vol. 19, No. 2, (Jan. 17, 2011), pp. 1470-1483. |
| Chen, L. et al., “Compact, low-loss and low-power 8x8 broadband silicon optical switch,” Optics Express, vol. 20, No. 17, pp. 18977-18985 (Aug. 13, 2012). |
| Chen, Q. et al., “A Universal method for constructing N-port non-blocking optical router based on 2x2 optical switch”, Progress In Electromagnetics Research Symposium Proceedings. Guangzhou. China. Aug. 25-28, 2014, pp. 357-361. |
| Cheng et al., “In-Plane Optical Absorption and Free Carrier Absorption in Graphene-on-Silicon Waveguides,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 20, No. 1, pp. 43-48, (Jan./Feb. 2014). |
| Chetlur et al., “cuDNN: Efficient primitives for deep learning”, arXiv preprint arXiv:1410.0759 (2014). 9 pages. |
| Childs, A. et al., “Spatial search by quantum walk”, Physical Review A, vol. 70, Issue 2, (Aug. 23, 2004), 12 pages. |
| Chung et al., “A monolithically integrated large-scale optical phased array in silicon-on-insulator CMOS”, IEEE Journal of Solid-State Circuits, vol. 53, No. 1, pp. 275-296 (Jan. 2018). |
| Cincotti, “Prospects on planar quantum computing”, Journal of Lightwave Technology, vol. 27, No. 24, pp. 5755-5766 (Dec. 15, 2009). |
| Clements et al., “Optimal design for universal multiport interferometers,” Optica, vol. 3, No. 12, pp. 1460-1465 (Dec. 2016). |
| Chinese Office Action and English Translation in Chinese Patent Application No. 201780043808.X dated Feb. 3, 2020, 12 pages. |
| Crespi, A. et al., “Integrated multimode interferometers with arbitrary designs for photonic boson sampling”, Nature Photonics, vol. 7, pp. 545-549 (May 26, 2013). |
| Crespi, et al., “Anderson localization of entangled photons in an integrated quantum walk”, Nature Photonics, vol. 7, Issue 4, pp. 322-328 (Apr. 2013). |
| Dai, D. et al., “Novel concept for ultracompact polarization splitter-rotator based on silicon nanowires”, Optics Express, vol. 19, No. 11, (May 23, 2011 ), pp. 10940-10949. |
| Di Giuseppe, G. et al., “Einstein-Podolsky-Rosen Spatial Entanglement in Ordered and Anderson Photonic Lattices”, Physical Review Letters, PRL 110, (Apr. 12, 2013), pp. 150503-1-150503-5. |
| Dunningham et al., “Efficient comparison of path-lengths using Fourier multiport devices”, Journal of Physics B: Atomic, Molecular and Optical Physics, vol. 39 (2006), pp. 1579-1586. |
| EP Extended European Search Report in European Patent Application No. 17807562.8, dated May 6, 2020, 14 pages. |
| EP Partial Supplemental European Search Report in European Patent Application No. 17807562.8 dated Jan. 17, 2020, 12 pages. |
| Esser et al., “Convolutional networks for fast, energy-efficient neuromorphic computing”, Proceedings of the National Academy of Sciences, vol. 113, No. 41, pp. 11441-11446 (Oct. 2016). |
| Farht et al., “Optical implementation of the Hopfield model”, Applied Optics, vol. 24, No. 10, pp. 1469-1475 (May 15, 1985). |
| Feinberg et al., “Making memristive neural network accelerators reliable”, 2018 IEEE International Symposium on High Performance Computer Architecture (HPCA), pp. 52-65 (2018). |
| Fushman, I. et al., “Controlled Phase Shifts with a Single Quantum Dot”, Science, vol. 320, (May 9, 2008), pp. 769-772. |
| George et al., “A programmable and configurable mixed-mode FPAA Soc”, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 24, No. 6, pp. 2253-2261 (Jun. 2016). |
| Gilmer et al., “Neural message passing for quantum chemistry”, arXiv preprint arXiv: 1704.01212 (2017). 14 pages. |
| Golub et al., “Calculating the singular values and pseudo-inverse of a matrix”, Journal of the Society for Industrial and Applied Mathematics Series B Numerical Analysis, vol. 2, No. 2, pp. 205-224 (1965). |
| Graves et al., “Hybrid computing using a neural network with dynamic external memory,” Nature, vol. 538, (Oct. 27, 2016), 21 pages. |
| Green, W. et al., “CMOS Integrated Silicon Nanophotonics: Enabling Technology for Exascale Computational System”, IBM Corporation, Invited Talk at SEMICON Dec. 1, 2010, Tokyo, (2010) 30 pages. |
| Grote et al., “First long-term application of squeezed states of light in a gravitational-wave observatory”, Physical Review Letters, PRL 110, 181101 (May 3, 2013). 5 pages. |
| Gruber et al., “Planar-integrated optical vector-matrix multiplier”, Applied Optics, vol. 39, No. 29, p. 5367-5373 (Oct. 10, 2000). |
| Gullans et al., “Single-Photon Nonlinear Optics with Graphene Plasmons”, Physical Review Letters, PRL 111, pp. 247401-1-247401-5 (Dec. 13, 2013) 7 pages. |
| Gunn, C., “CMOS photonics for high-speed interconnects”, IEEE Micro, (Mar.-Apr. 2006), pp. 58-66. |
| Haffner et al., “Low-loss plasmon-assisted electro-optic modulator”, Nature, vol. 556, pp. 483-486 (2018). |
| Halasz et al., “Phase diagram of QCD”, Physical Review D, vol. 5 8, pp. 096007-1-096007-11, (1998). |
| Hamerly et al., “Scaling advantages of all-to-all connectivity in physical annealers: the Coherent Ising Machine vs. DWave 2000Q”, arXiv: 1805.05217, (May 2018), 17 pages. |
| Harris et al. “Integrated source of spectrally filtered correlated photons for large-scale quantum photonic systems”, Physical Review X, vol. 4, Issue 4, pp. 041047-1-041047-10 (2014). |
| Harris et al., “Bosonic transport simulations in a large-scale programmable nanophotonic processor”, arXiv: 1507,03406v2, 8 pages (2015). |
| Harris et al., “Efficient, compact and low loss thermos-optic phase shifter in silicon”, Optics Express, vol. 22, No. 9, pp. 10487-10493 (2014). |
| Hinton et al., “Reducing the dimensionality of data with neural networks”, Science, vol. 313, pp. 504-507 (Jul. 28, 2006). |
| Hochberg, M. et al., “Silicon Photonics: The Next Fabless Semiconductor Industry”, IEEE Solid-State Circuits Magazine, pp. 48-58 (Winter 2013). |
| Honerkamp-Smith et al., “An introduction to critical points for biophysicists; observations of compositional heterogeneity in lipid membranes”, Biochimica et Biophysica Acta 1788, pp. 53-63, (2009). |
| Hong, C. K. et al., “Measurement of subpicosecond time intervals between two photons by interference”, Physical Review Letters, vol. 59, No. 18, (Nov. 2, 1987), pp. 2044-2046. |
| Hopefield et al., “Neural computation of decisions in optimization problems”, Biological Cybernetics, vol. 52, No. 3, pp. 141-152 (1985). |
| Hopefield, “Neural networks and physical systems with emergent collective computational abilities”, Proceedings of the National Academy of Sciences of the United States of America, vol. 79, pp. 2554-2558 (Apr. 1982). |
| Horowitz, M., “Computing's energy problem (and what we can do about it)”, 2014 IEEE International Solid-State Circuits Conference, Digest of Technical Papers, ISSCC 2014, Session 1, Plenary / 1.1 (Feb. 2014). |
| Horst, F. et al., “Cascaded Mach-Zehnder wavelength filters in silicon photonics for low loss and flat pass-band WDM (de-)multiplexing”, Optics Express, vol. 21, No. 10, pp. 11652-11658 (2013). |
| Humphreys, P. C. et al., “Linear Optical Quantum Computing in a Single Spatial Mode”, arXiv: 1305.3592, (Nov. 21, 2014), 7 pages. |
| Inagaki et al., Large-scale Ising spin network based on degenerate optical parametric oscillators, Nature Photonics. vol. 10, pp. 415-419 (Jun. 2016). |
| Isichenko, “Percolation, statistical topography, and transport in random media”, Reviews of Modern Physics, vol. 64, No. 4, pp. 961-1043, (Oct. 1992). |
| Ising, “Beitrag zur Theorie des Ferromagnetismus,” Z. Phys., pp. 253-258 (Dec. 1924). |
| Jaekel et al., “Quantum limits in interferometric measurements”, Europhysics Letters, vol. 13, No. 4, pp. 301-306 (Oct. 1990). |
| Jalali, B. et al., “Silicon Photonics”, Journal of Lightwave Technology, vol. 24, No. 12, pp. 4600-4615 (Dec. 2006). |
| Jia et al., “Caffe: Convolutional architecture for fast feature embedding”, Proceedings of the 22nd ACM International Conference on Multimedia (Multimedia 2014), Orlando, Florida, pp. 675-678 URL http://doi.acm. org/1 O.1145/264 7868 .2654889 (2014). |
| Jiang, L. et al., “A planar ion trapping microdevice with integrated waveguides for optical detection”, Optics Express, vol. 19, No. 4, pp. 3037-3043 (Feb. 2011). |
| Jonsson, “An empirical approach to finding energy efficient ADC architectures”, 2011 International Workshop on ADC Modelling, Testing and Data Converter Analysis and Design and IEEE 2011 ADC Forum Jun. 30-Jul. 1, 2011, Orvieto, Italy (2011). |
| Jouppi et al. “In-datacenter performance analysis of a tensor processing unit”, 2017 ACM/IEEE 44th Annual International Symposium on Computer Architecture (ISCA) (Jun. 24-28, 2017). |
| Kahn et al., “Communications expands its space”, Nature Photonics, vol. 11, pp. 5-8 (Jan. 2017). |
| Kardar et al., “Dynamic Scaling of Growing Interfaces”, Physical Review Letters, vol. 56, No. 9, pp. 889-892 (Mar. 3, 1986). |
| Karpathy, A., “CS231n: Convolutional Neural Networks for Visual Recognition”, Class notes. Jan. 2018, http://cs231 n.github.io/. Accessed Oct. 31, 2018. 2 pages. |
| Keckler et al., “GPUs and the future of parallel computing”, IEEE Micro, pp. 7-17 (2011). |
| Kieling, K. et al., “On photonic Controlled Phase Gates”, New Journal of Physics, vol. 12, (2010), 9 pages. |
| Kilper et al., “Optical networks come of age”, Optics and Photonics News, vol. 25, pp. 50-57, (Sep. 2014). |
| Kim et al., “A functional hybrid memristor crossbar-array/CMOS system for data storage and neuromorphic applications” Nano Letters, vol. 12, pp. 389-395 (2012). |
| Kirkpatrick et al., “Optimization by simulated annealing”, Science, vol. 220, No. 4598, pp. 671-680 (May 13, 1983). |
| Knill et al., “The Bayesian brain: the role of uncertainty in neural coding and computation”, Trends in Neurosciences, vol. 27, No. 12, pp. 712-719 (Dec. 2004). |
| Knill, E. et al., “A scheme for efficient quantum computation with linear optics”, Nature, vol. 409, pp. 46-52 (Jan. 4, 2001). |
| Knill, E., “Quantum computing with realistically noisy devices”, Nature, vol. 434, pp. 39-44 (Mar. 3, 2005). |
| Kok et al. “Linear optical quantum computing with photonic qubits”, Reviews of Modern Physics, vol. 79, pp. 135-174 (2007). |
| Koos et al., “Silicon-organic hybrid (SOH) and plasmonic-organic hybrid (POH) integration”, Journal of Lightwave Technology, vol. 34, No. 2, pp. 256-268 (2016). |
| Krizhevsky et al., “ImageNet classification with deep convolutional neural networks”, NIPS'12: Proceedings of the 25th International Conference on Neural Information Processing Systems, pp. 1097-1105, (Dec. 2012). |
| Kucherenko, S et al., “Application of Deterministic Low-Discrepancy Sequences in Global Optimization”, Computational Optimization and Applications, vol. 30, pp. 297-318, (2005). |
| Kwack, M-J et al., “Monolithic InP strictly non-blocking 8x8 switch for high-speed WDM optical interconnection”, Optics Express, vol. 20, No. 27, pp. 28734-28741, (Dec. 2012). |
| Lahini, Y et al., “Anderson Localization and Nonlinearity in One Dimensional Disordered Photonic Lattices”, Physical Review Letters, vol. 100, (Feb. 2008), 4 pages. |
| Lahini, Y et al., “Quantum Correlations in Two-Particle Anderson Localization”, Physical Review Letters, Phys. Rev. Lett. 105, pp. 163905-1-163905-4 (Oct. 15, 2010). |
| Laing, A et al., “High-fidelity operation of quantum photonic circuits”, Applied Physics Letters, vol. 97, No. 21, (Apr. 2010), 5 pages. |
| Landauer, “Irreversibility and heat generation in the computing process”, IBM Journal of Research and Development, pp. 183-191, (Jul. 1961). |
| Lanyon, B. P. et al., “Towards quantum chemistry on a quantum computer”, Nature Chemistry, vol. 2, pp. 106-111, (May 8, 2009). |
| Lawson et al., “Basic linear algebra subprograms for Fortran usage”, ACM Transactions on Mathematical Software, vol. 5, No. 3, pp. 308-323, (Sep. 1979). |
| Lecun et al., “Deep learning”, Nature, vol. 521, pp. 436-444, (May 2015). |
| Lecun et al., “Gradient-based learning applied to document recognition”, Proceedings of the IEEE, vol. 86, No. 11, pp. 2278-2324 (Nov. 1998). |
| Levi, L. et al., “Hyper-transport of light and stochastic acceleration by evolving disorder”, Nature Physics, vol. 8, pp. 912-917 (Dec. 2012). |
| Li et al., “Efficient and self-adaptive in-situ learning in multilayer memristor neural networks”, Nature Communications, vol. 9, No. 1, (Jun. 2018), 8 pages. |
| Little, “The existence of persistent states in the brain”, Mathematical Biosciences, vol. 19, Nos. 101-120, pp. 145-165 (1974). |
| Lu et al., “16 x 16 non-blocking silicon optical switch based on electro-optic Mach-Zehnder interferometers”, Optics Express, vol. 24, No. 9, pp. 9295-9307, (2016). |
| Ma et al., “Optical switching technology comparison: optical MEMS vs. Other technologies”, IEEE Optical Communications, vol. 41, No. 11, pp. S16-S23 (2003). |
| Macready et al., “Criticality and Parallelism in Combinatorial Optimization”, Science, vol. 271, pp. 56-59, (Jan. 1996). |
| Marandi et al., “Network of time-multiplexed optical parametric oscillators as a coherent Ising machine”, Nature Photonics, vol. 8, pp. 937-942 (Dec. 2014). |
| Martin-Lopez, E. et al., “Experimental realization of Shor's quantum factoring algorithm using qubit recycling”, Nature Photonics, vol. 6, (Oct. 24, 2012), 7 pages. |
| McMahon et al., “A fully programmable 100-spin coherent Ising machine with all-to-all connections.,” Science, vol. 354, Issue 6312, pp. 614-617, (Nov. 2016). |
| Mead, “Neuromorphic electronic systems”, Proceedings of the IEEE, vol. 78, No. 10, pp. 1629-1636 (Oct. 1990). |
| Migdall, A. L. et al., “Tailoring single and multiphoton probabilities of a single photon on-demand source”,Physical Review A, vol. 66, pp. 053805-1-053805-4 (2002). |
| Mikkelsen, J.C. et al., “Dimensional variation tolerant silicon-on-insulator directional couplers”, Optics Express, vol. 22, No. 3, pp. 3145-3150 (2014). |
| Miller, “Are optical transistors the logical next step?”, Nature Photonics, vol. 4, pp. 3-5 (2010). |
| Miller, “Attojoule optoelectronics for low-energy information processing and communications”, Journal of Lightwave Technology, vol. 35, No. 3, pp. 346-396 (Feb. 2017). |
| Miller, D. A. B., “Perfect optics with imperfect components”, Optica, vol. 2, No. 8, pp. 747-750 (Aug. 2015). |
| Miller, D. A. B., “Reconfigurable add-drop multiplexer for spatial modes”, Optics Express, vol. 21, No. 17, pp. 20220-20229, (Aug. 2013). |
| Miller, D. A. B., “Self-aligning universal beam coupler”, Optics Express, vol. 21, No. 5 (Aug. 2013), 6 pages. |
| Miller, D. A. B., “Self-configuring universal linear optical component [invited]”, Photonics Research, vol. 1, No. 1, pp. 1-15 (Jun. 2013). |
| Miller, “Energy consumption in optical modulators for interconnects”, Optics Express, vol. 20, No. S2, pp. A293-A308 (2012). |
| Misra et al., “Artificial neural networks in hardware: A survey of two decades of progress”, Neurocomputing, vol. 74, pp. 239-255 (2010). |
| Mohseni, M. et al., “Environment-assisted quantum walks in energy transfer of photosynthetic complexes”, The Journal of Chemical Physics, vol. 129, (2008), 8 pages. |
| Moore, “Cramming more components onto integrated circuits”, Proceedings of the IEEE, vol. 86, No. 1, pp. 82-85 Jan. 1998). |
| Mower et al., “High-fidelity quantum state evolution in imperfect photonic integrated circuits”, Physical Review A, vol. 92, No. 3, pp. 032322-1-032322-7 (2015). |
| Mower, J. et al., “Efficient generation of single and entangled photons on a silicon photonic integrated chip”, Physical Review A, vol. 84, (2011), 8 pages. |
| Nagamatsu et al., “A 15 NS 32X32-bit CMOS multiplier with an improved parallel structure”, Custom Integrated Circuits Conference, Proceedings of the IEEE 1989, (1989), 4 pages. |
| Najafi, F. et al., “On-Chip Detection of Entangled Photons by Scalable Integration of Single-Photon Detectors”, arXiv:1405,4244 [physics.optics](May 16, 2014), 27 pages. |
| Nozaki et al., “Sub-femtojoule all-optical switching using a photonic-crystal nanocavity”, Nature Photonics, vol. 4, pp. 477-483 (2010). |
| O'Brien, J. L. et al., “Demonstration of an all-optical quantum controlled-NOT gate”, Nature, vol. 426, (2003), 5 pages. |
| Onsager, “Crystal Statistics. I. A Two-Dimensional Model with an Order-Disorder Transition”, Physical Review, vol. 65, Nos. 3 and 4, pp. 117-149 (Feb. 1944). |
| Orcutt, J. S et al., “Nanophotonic integration in state-of-the-art CMOS foundries”, Optics Express, vol. 19, No. 3, pp. 2335-2346 (2011). |
| Patel et al., “Silicon Photonic Segmented Modulator-Based Electro-Optic DAC for 100 GB/s PAM-4 Generation”, IEEE Photonics Technology Letters, vol. 27, No. 23, pp. 2433-2436 (Dec. 1, 2015). |
| PCT International Preliminary Report on Patentability for International Application No. PCT/US2019/035403 dated Dec. 8, 2020, 10 pages. |
| PCT International Search Report and Written Opinion in PCT/US2018/041640 dated Nov. 7, 2018, 10 pages. |
| PCT International Search Report and Written Opinion mailed Sep. 28, 2017 from International Application No. PCT/US2017/035668, 19 pages. |
| Pelissetto et al., “Critical phenomena and renormalization-group theory”, Physics Reports, vol. 368, pp. 549-727, (2002). |
| Peng, “Implementation of AlexNet with Tensorflow”, https://github.com/ykpengba/AlexNet-A- Practical-Implementation (2018), Accessed Dec. 3, 2018, 2 pages. |
| Peretto, “Collective properties of neural networks: A statistical physics approach”, Biological Cybernetics, vol. 50, pp. 51-62 (1984). |
| Pernice, W. et al., “High-speed and high-efficiency travelling wave single-photon detectors embedded in nanophotonic circuits”, Nature Communications, vol. 3, (2012), 23 pages. |
| Peruzzo, A., et al., “Quantum walk of correlated particles”, Science, vol. 329, Issue 5998 (2010), 8 pages. |
| Politi, A. et al., “Integrated Quantum Photonics”, IEEE Journal of Selected Topics in Quantum Electronics, vol. 5, Issue 6, (2009), 12 pages. |
| Politi, A. et al., “Silica-on-Silicon Waveguide Quantum Circuits”, Science, vol. 320, (2008), 5 pages. |
| Poon et al., “Neuromorphic silicon neurons and large-scale neural networks: challenges and opportunities”, Frontiers in Neuroscience, vol. 5, Article 108, 3 pages (2011). |
| Prucnal et al., “Recent progress in semiconductor excitable lasers for photonic spike processing”, Advances in Optics and Photonics, vol. 8, No. 2, pp. 228-299 (Jun. 2016). |
| Psaltis et al., “Holography in artificial neural networks”, Nature, vol. 343, pp. 325-330 (Jan. 1990). |
| Qiao et al., “16x16 non-blocking silicon electro-optic switch based on Mach Zehnder interferometers”, 2016 Optical Fiber Communications Conference and Exhibition (OFC), (2016), 3 pages. |
| Ralph, T. C. et al., “Linear optical controlled-NOT gate in the coincidence basis”, Physical Review A, vol. 65, pp. 062324-1-062324-5 (2002). |
| Ramanitra et al., “Scalable and multi-service passive optical access infrastructure using variable optical splitters”, Optical Fiber Communication Conference. Optical Society of America, (2006), 3 pages. |
| Raussendorf, R. et al., “A one-way quantum computer”, Physical Review Letters, vol. 86, No. 22, pp. 5188-5191 (2001). |
| Rechtsman et al., “Photonic floquet topological insulators”, Optical Society of America, Technical Digest, 2 pages (2013). |
| Reck et al., “Experimental realization of any discrete unitary operator,” Physical Review Letters, vol. 73, No. 1, pp. 58-61 (1994). |
| Reed, G. T. et al., “Silicon optical modulators”, Nature Photonics, vol. 4, pp. 518-526 (2010). |
| Rendl et al., “Solving Max-Cut to optimality by intersecting semidefinite and polyhedral relaxations”, Mathematical Programming, vol. 121, pp. 307-335, (2010). |
| Rios et al., “Integrated all-photonic non-volatile multilevel memory”, Nature Photonics, vol. 9, pp. 725-732 (Nov. 2015). |
| Rogalski, “Progress in focal plane array technologies”, Progress in Quantum Electronics, vol. 36, pp. 342-473 (2012). |
| Rohit, A. et al., “8x8 space and wavelength selective cross-connect for simultaneous dynamic multi-wavelength routing”, Optical Fiber Communication Conference/National Fiber Optic Engineers Conference 2013, (2013), 3 pages. |
| Rosenblatt, “The perceptron: a probabilistic model for information storage and organization in the brain”, Psychological Review, vol. 65, No. 6, (1958) 23 pages. |
| Russakovsky et al. “ImageNet Large Scale Visual Recognition Challenge”, International Journal of Computer Vision (IJCV), vol. 115, pp. 211-252 (2015). |
| Salandrino, A et al., “Analysis of a three-core adiabatic directional coupler”, Optics Communications, vol. 282, pp. 4524-4526 (2009). |
| Schaeff et al., “Scalable fiber integrated source for higher-dimensional path-entangled photonic quNits”, Optics Express, vol. 20, No. 15 pp. 16145-16153 (2012). |
| Schell et al., “DAC-free Generation of M-QAM Signals with InP Segmented Mach-Zehnder Modulators”, Optical Fiber Communication Conference, OSA Technical Digest (online) (Optical Society of America, 2017), paper W4G.4. |
| Schirmer et al., “Nonlinear mirror based on two-photon absorption”, Journal of Optical Society of America B, vol. 14, Issue 11, pp. 2865-2868 (1997). |
| Schmidhuber, J., “Deep learning in neural networks: An overview”, Neural Networks, vol. 61, pp. 85-117 (2015). |
| Schreiber, A. et al., “Decoherence and Disorder in Quantum Walks: From Ballistic Spread to Localization”, Physical Review Letters, vol. 106, (2011), 5 pages. |
| Schwartz, T. et al., “Transport and Anderson localization in disordered two-dimensional photonic lattices”, Nature, vol. 446, pp. 52-55 (Mar. 1, 2007). |
| Selden, “Pulse transmission through a saturable absorber”, British Journal of Applied Physics, vol. 18, (1967) 7 pages. |
| Shafiee et al., “ISAAC: A convolutional neural network accelerator with in-situ analog arithmetic in crossbars,” 2016 ACM/IEEE 43rd Annual International Symposium on Computer Architecture, (2016) 13 pages. |
| Shoji, Y. et al., “Low-crosstalk 2x2 thermo-optic switch with silicon wire waveguides”, Optics Express, vol. 18, No. 9, pp. 9071-9075, (Apr. 2010). |
| Silver et al. “Mastering chess and shogi by self-play with a general reinforcement learning algorithm”, arXiv preprint arXiv:1712,01815 (2017) 19 pages. |
| Silver et al., “Mastering the game of GO with deep neural networks and tree search”, Nature, vol. 529, pp. 484-489 (2016). |
| Silverstone, J. et al., “On-chip quantum interference between silicon photon-pair sources”, Nature Photonics, vol. 8, pp. 104-108 (2014). |
| Smith et al., “Phase-controlled integrated photonic quantum circuits”, Optics Express, vol. 17, No. 16, pp. 13516-13525 (2009). |
| Soljacic et al., “Optimal bistable switching in nonlinear photonic crystals”, Physical Review E, vol. 66, pp. 055601-1-055601-4 (2002). |
| Solli et al., “Analog optical computing”, Nature Photonics, vol. 9, pp. 704-706 (Nov. 2015). |
| Spring, J. B et al., “Boson sampling on a photonic chip”, Science, vol. 339, (2013), 24 pages. |
| Srinivasan et al., “56 GB/s germanium waveguide electro-absorption modulator”, Journal of Lightwave Technology, vol. 34, No. 2, pp. 419-424 (Jan. 15, 2016). |
| Steinkraus et al., “Using GPUs for machine learning algorithms”, Proceedings of the 2005 Eight International Conference on Document Analysis and Recognition (ICDAR'05), pp. 1115-1120 (2005). |
| Suda et al., “Quantum interference of photons in simple networks”, Quantum Information Processing, vol. 12, pp. 1915-1945 (2013). |
| Sun et al., “Large-scale nanophotonic phased array”, Nature, vol. 493, pp. 195-199 (2013). |
| Sun et al., “Single-chip microprocessor that communicates directly using light”, Nature, vol. 528, pp. 534-538 (2015). |
| Suzuki, K. et al., “Ultra-compact 8x8 strictly-non-blocking Si-wire PILOSS switch”, Optics Express, vol. 22, No. 4, pp. 3887-3894 (2014). |
| Sze et al., “Efficient processing of deep neural networks: A tutorial and survey”, Proceedings of the IEEE, vol. 105, No. 12, pp. 2295-2329 (2017). |
| Tabia, “Experimental scheme for qubit and qutrit symmetric informationally complete positive operator-valued measurements using multiport devices”, Physical Review A, vol. 86, pp. 062107-1-062107-8 (2012). |
| Tait et al., “Broadcast and weight: An integrated network for scalable photonic spike processing”, Journal of Lightwave Technology, vol. 32, No. 21, p. 3427-3439, (2014). |
| Tait et al., “Photonic Neuromorphic Signal Processing and Computing”, Nanophotonic Information Physics, pp. 183-222, Springer, Berlin, Heidelberg, (2014). |
| Tait et al., “Neuromorphic photonic networks using silicon photonic weight banks”, Scientific Reports, vol. 7, Article No. 7430 (2017), 10 pages. |
| Tanabe et al., “Fast bistable all-optical switch and memory on a silicon photonic crystal on-chip”, Optics Letters, vol. 30, No. 19, pp. 2575-2577 (Oct. 1, 2005). |
| Tanizawa, K. et al., “Ultra-compact 32x32 strictly-non-blocking Si-wire optical switch with fan-out LGA interposer”, Optics Express, vol. 23, No. 13, pp. 17599-17606 (2015). |
| Thompson, M. G. et al., “Integrated waveguide circuits for optical quantum computing”, IET Circuits, Devices & Systems, vol. 5, Issue 2, pp. 94-102 (2011). |
| Timurdogan et al., “An ultralow power athermal silicon modulator”, Nature Communications 5, Article No. 4008 (2014) 11 pages. |
| Taiwan Office Action for Taiwan Application No. TW 109109209, dated Mar. 4, 2021, 9 pages (With English Translation). |
| Vandoorne et al., “Experimental demonstration of reservoir computing on a silicon photonics chip”, Nature Communications, vol. 5, (2014) 6 pages. |
| Vazquez et al., “Optical NP problem solver on laser-written waveguide platform”, Optics Express, vol. 26, No. 2, pp. 702-710 (Jan. 2018). |
| Vivien et al., “Zero-bias 40Gbit/s germanium waveguide photodetector on silicon”, Optics Express, vol. 20, No. 2, pp. 1096-1101 (2012). |
| Wang et al., “Coherent Ising machine based on degenerate optical parametric oscillators” Physical Review A, vol. 88, p. 063853-1-063853-9 (2013). |
| Wang et al., “Deep learning for identifying metastatic breast cancer”, arXiv: 1606,05718 (2016) 6 pages. |
| Werbos, “Beyond regression: New tools for prediction and analysis in the behavioral sciences”, Ph.D, dissertation, Harvard University (1974) 454 pages. |
| Whitfield, J. D. et al., “Simulation of electronic structure Hamiltonians using quantum computers”, Molecular Physics, vol. 109, pp. 735-750 (2010). |
| Wu et al., “A 20Gb/s NRZ/PAM-4 IV Transmitter in 40nm CMOS Driving a Si-Photonic Modulator in 0.13 μm CMOS”, IEEE International Solid-State Circuits Conference, Session 7, Optical Transceivers and Silicon Photonics, 7.7, pp. 128-130 (2013). |
| Wu et al., “An optical fiber network oracle for NP-complete problems”, Light: Science & Applications, vol. 3, (2014). |
| Xia, F., et al., “Mode conversion losses in silicon-on-insulator photonic wire based racetrack resonators”, Optics Express, vol. 14, No. 9, pp. 3872-3886 (2006). |
| Xu et al., “Experimental observations of bistability and instability in a two-dimensional nonlinear optical superlattice”, Physical Review Letters, vol. 71, No. 24, pp. 3959-3962 (Dec. 13, 1993). |
| Yang, M. et al., “Non-Blocking 4x4 Electro-Optic Silicon Switch for On-Chip Photonic Networks”, Optics Express, vol. 19, No. 1, pp. 47-54 (2010). |
| Yao et al., “Serial-parallel multipliers”, Signals, Systems and Computers, Conference Record of The Twenty-Seventh Asilomar Conference, pp. 359-363 (1993). |
| Young et al., “Recent trends in deep learning based natural language processing”, IEEE Computational Intelligence Magazine, pp. 55-75 (Aug. 2018). |
| Zhou, X.-Q., et al., “Calculating Unknown Eigenvalues with a Quantum Algorithm”, Nature Photonics 7, pp. 223-228 (2013). |
| The Taiwan Office Action for Taiwan Application No. TW 10914293 6 dated Jul. 29, 2021 (Waiting for English Translation). |
| U.S. Appl. No. 63/048,439, filed Jul. 6, 2020, Meng et al. |
| U.S. Appl. No. 63/061,995, filed Aug. 6, 2020, Meng et al. |
| U.S. Appl. No. 63/123,338, filed Dec. 9, 2020, Wu et al. |
| The Notice of Allowance and Search Report issued by the Taiwan Patent Office for Application No. TW 111105772, dated May 4, 2022 (with English Translation). |
| The Taiwan IPO Examination Report and Search Report for Application No. 110115269 dated Jun. 7, 2022 (with English Translation). |
| Number | Date | Country | |
|---|---|---|---|
| 20210341765 A1 | Nov 2021 | US |
| Number | Date | Country | |
|---|---|---|---|
| 63017211 | Apr 2020 | US |
