Patent Application
20190341115
The present disclosure relates generally to memory and in particular, in one or more embodiments, the present disclosure relates to methods and apparatus for facilitating neural network operations in a memory device.
Memory devices are typically provided as internal, semiconductor, integrated circuits in computers or other electronic systems. There are many different types of memory including random-access memory (RAM), read only memory (ROM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), and flash memory.
Flash memory devices have developed into a popular source of non-volatile memory for a wide range of electronic applications. Flash memory devices typically use a one-transistor memory cell that allows for high memory densities, high reliability, and low power consumption. Changes in threshold voltage of the cells, through programming of a charge storage structure, such as floating gates or trapping layers or other physical phenomena, determine the data state of each cell. Flash memory devices are commonly used in electronic systems, such as personal computers, personal digital assistants (PDAs), digital cameras, digital media players, digital recorders, games, appliances, vehicles, wireless devices, cellular telephones, and removable memory modules, and the uses for flash memory continue to expand.
Flash memory typically utilizes one of two basic architectures known as NOR flash and NAND flash. The designation is derived from the logic used to read the devices. In NOR flash architecture, a logical column of memory cells is coupled in parallel with each memory cell coupled to a data line, such as those typically referred to as digit (e.g., bit) lines. In NAND flash architecture, a column of memory cells is coupled in series with only the first memory cell of the column coupled to a bit line.
Neural networks are networks which process information by modeling a network of neurons, such as neurons in a human brain, to process information (e.g., stimuli) which has been sensed in a particular environment. Similar to a human brain, neural networks typically comprise multiple neuron models to process information. The demand for improved operating characteristics of neural networks continues to increase. Such desirable neural network operating characteristics improvements are increased speed, capacity and processing power of neural networks, for example.
For the reasons stated above, and for other reasons stated below which will become apparent to those skilled in the art upon reading and understanding the present specification, there is a need in the art for methods and devices for improving operating characteristics of neural networks.
In the following detailed description of the invention, reference is made to the accompanying drawings that form a part hereof, and in which is shown, by way of illustration, specific embodiments. In the drawings, like numerals describe substantially similar components throughout the several views. Other embodiments may be utilized and structural, logical, and electrical changes may be made without departing from the scope of the present disclosure. The following detailed description is, therefore, not to be taken in a limiting sense.
The memory array 100 shown in
Each series string 104, 106 of memory cells is coupled to a source 112 by a source select gate 114, 116 (e.g., transistor) and to an individual even or odd bit line BL_E, BL_O by a drain select gate 118, 120 (e.g., transistor). The source select gates 114, 116 are controlled by a source select gate control line SGS 122 coupled to their control gates. The drain select gates 118, 120 are controlled by a drain select gate control line SGD 124.
In a typical programming of the memory array 100, each memory cell is individually programmed as either a single level cell (SLC) or a multiple level cell (MLC). The cell's threshold voltage (Vt) can be used as an indication of the data stored in the cell. For example, in an SLC, a Vt of 2.5V might indicate a first data state while a Vt of −0.5V might indicate a second data state. An MLC uses multiple Vt ranges that each indicates a different data state. Multiple level cells can take advantage of the analog nature of a traditional charge storage cell by assigning a bit pattern to a specific Vt range. Single level memory cells permit the storage of a single binary digit (e.g., bit) of data on each memory cell. Meanwhile, multiple level cells permit the storage of two or more binary digits per cell (e.g., 2, 3, 4, 5 bits), depending on the quantity of Vt ranges assigned to the cell and the stability of the assigned Vt ranges during the lifetime operation of the memory cell. By way of example, one bit (e.g., 1 or 0) may be represented by two Vt ranges, two bits by four ranges, three bits by eight ranges, etc
Neural networks implement methods for interpreting particular types of information (e.g., data), such as information obtained in response to sensing a physical phenomenon (e.g., sensing stimuli) in a particular environment, and in response generating one or more outputs. A common goal of neural networks is to mimic a system, such as the human brain, so as to learn to interpret and respond to particular stimuli. For example, human beings have an ability to observe and identify an object even without having an exact match in their memory of what is being observed. A person might be able to recognize another person even though one or more traits of the other person have changed. For example, a person might still be able to recognize another person even though the other person has changed their hair color, grown or removed facial hair, etc.
Electronic systems typically are proficient at confirming an exact match between something sensed by the system, such as with an electronic camera, and something stored in its memory. However, electronic systems are less proficient at identifying and interpreting an object that might be similar, yet somewhat different, than what the system might have stored in its memory. Continuing with the previous example, an electronic system might not be able to identify a person whose image is stored in the system memory if that person has changed their hair color or added/removed facial hair, etc.
A typical structure of neural networks is to mimic networks of neurons, such as found in a human brain, by utilizing small building blocks (e.g., neuron models) to attempt to emulate the neurons of a neural network.
Various functions might be utilized for the function 208. For example, function 208 might comprise a function to determine if the summed value 206 is above or below a particular threshold level. Such a threshold function might generate a logic high output potential on the output 220 if the summed inputs 206 are equal to or greater than the particular threshold level and might generate a logic low potential on the output 220 if the summed inputs 206 are below the particular threshold level, for example. According to one or more embodiments of the present disclosure, the function 208 might comprise a sigmoid function, wherein the sigmoid function might be expressed as f(x)=1/(1+e−λx), where λ might comprise some constant value. Thus, the function 208 might comprise a non-linear function. The generated output value at the output 220 can be applied to one or more additional neuron models (e.g., such as to inputs 202 of different neuron models) which comprise a neural network of neuron models according to various embodiments of the present disclosure.
The input stage 304 comprising initial input signals (e.g., VIN1, VIN2) to the neural network 300 may have been supplied by additional circuitry of an electronic system incorporating the neural network. Each input node 312 might comprise circuitry such as registers and/or driver circuits, for example. The output stage 310 is configured to transfer neural network output values to additional circuitry of the electronic system. Output nodes 314 might comprise circuitry such as registers and/or driver circuits, for example.
Memory devices and methods according to various embodiments of the present disclosure include operating memory cells of a memory array as one or more neuron models such as to facilitate a neural network in the memory device. The memory array might comprise a NAND configured array of charge storage (e.g., Flash) memory cells, for example.
According to various embodiments of the present disclosure, a neuron model as facilitated in a NAND array of memory cells may comprise one or more concurrently selected memory cells of one or more strings of memory cells coupled to a common bit line. The number of inputs of a neuron model might equal the number of selected strings comprising a selected memory cell which are coupled to the same bit line. For example, one or more two input neuron models may be described by way of reference to
A memory cell of each of one or more strings of memory cells coupled to the same bit line can be selected to collectively form a portion of a neuron model. Thus, the neuron model might comprise a selected memory cell from one or more strings of memory cells, a bit line coupled to each of the selected strings of memory cells and the sense circuitry coupled to the bit line. For example a neuron model according to various embodiments of the present disclosure might comprise selected memory cells 4121 and 4122, bit line 408 and the sense circuitry 430. A different neuron model might comprise selected memory cells 4141 and 4142, the bit line 408 and sense circuitry 430, for example.
The operation of a neuron model incorporating one or more memory cells of one or more strings of memory cells might be further described by way of example. Word lines of a first sector of memory, such as Sector J 420, can be driven independently from word lines of a second sector, such as Sector J+1 422. Word lines of sectors 420, 422 are driven by word line drivers (not shown) coupled to the word lines. To operate the memory array 400 as a first neuron model, a memory cell from string 402 and a memory cell from string 404 are selected. For example, memory cells 4101 and 4102 might comprise the memory cells of a first neuron model having two inputs. It should be noted that each of the memory cells of string 402 and/or each of the memory cells of string 404 may have been programmed to a respective one of a number of different data states during a previous operation of the memory device.
The word lines 4201 and 4202 coupled to the selected memory cells 4101, 4102 might each be biased to a first potential and to a second potential, respectively. The first potential and the second potential comprise input potentials for the particular neuron model presently operating. The input potentials might comprise inputs received by a memory device as initial inputs for the neural network to act upon. Alternatively, the input potentials may comprise output signals generated previously by other neuron models of the memory device, for example. The signals (e.g., input potentials) applied to the word lines coupled to the selected memory cells might have been subject to some amount of signal conditioning prior to being applied to the word lines. The remaining word lines of each string (e.g., other than those coupled to selected memory cells) might be biased at a potential (e.g., Vpass potential) sufficient to cause the coupled memory cells to be activated (e.g., conduct) regardless of their respective data states.
During a sense operation, BL 408 might be pre-charged to a particular level prior to concurrently applying the first and second input potentials to the two selected memory cells 4101, 4102 comprising the first neuron model. As discussed above, unselected memory cells of strings 402 and 404 are operated in a pass through mode to conduct regardless of any data state they might be in. Thus, the amount of current present on the bit line BL 408 during the sense operation is dependent upon the data states of the two selected memory cells and the potentials (e.g., input potentials) applied to their respective word lines. For example, the bit line current on BL 408 comprises a sum of the current, if any, flowing through each string of memory cells responsive to the conductive state of the selected memory cells of strings 402 and 404. As discussed above, unselected memory cells of strings 402 and 404 are operated in a pass through mode to conduct regardless of any data state they might be in.
Sense circuitry 430 is configured to respond to the bit line current and generate an output responsive to the bit line current. The generated output might comprise various types of signals. The output signal might comprise a particular voltage level. The generated output signal might comprise a digital representation (e.g., one or more bits) representative of the current sensed in the bit line BL 408, for example. The generated output signal might then be applied to an input of one or more different neuron models and/or to an output of the neural network, for example.
A second neuron model (e.g., different neuron model than the neuron model comprising memory cells 4101, 4102 described above) might be facilitated utilizing the same strings of memory cells 402, 404 coupled to the same bit line BL 408 shown in
The number of inputs (e.g., such as up to INPUT_N 2033 shown in
The memory device illustrated in
A method of facilitating a neural network in a memory device according to one or more embodiments of the present disclosure can be described by way of reference to
The first level of neuron models of
During a first sense (e.g., read) operation of the neural network illustrated by
Upon completion of the first sense operation of the neuron models 540 and 542, the word line register 532 receives the sensed data values transferred over the output signal lines 520, 522 and stores them in the word line register. A second sense operation is then performed on a second level of neuron models. The second level of neuron models are represented (e.g., by dashed lines) by neuron models 544, 546 shown in
During the second sense operation of the neural network illustrated by
A third sense operation is illustrated by way of reference to
A fourth sense operation is illustrated by way of reference to
Upon completion of the fourth sense operation of the neural network, the output data values, such as generated by sense circuitry 516, 518 during the fourth sense operation, might be transferred from the sense circuitry registers to the word line register 532 as discussed above. The word line register 532 can transfer the data values from the word line register over the bidirectional interface 538 to another portion of the memory device (not shown) such as control logic or to a memory access device (e.g., processor) coupled to the memory device, for example. According to additional embodiments, the output data values generated responsive to the fourth sense operation might be made available as input values to another neural network within the memory device, such as located in different sectors, pages or blocks of memory, for example. Thus, the output data values of a first neural network might comprise the input values for a second neural network within the memory device according to various embodiments of the present disclosure.
Although the example discussed with respect to
The operating characteristics of a neural network according to various embodiments of the present disclosure might be at least partially defined by programming the memory cells used in neuron models comprising the neural network to various data states. The respective data state (e.g., threshold voltage Vt) of individual memory cells can facilitate a weighting function respective to inputs applied to the memory cells. For example, a higher Vt of a memory cell will require a higher input potential be applied to activate the memory cell. These data states might be determined in various ways. For example, the data states might be determined by applying particular stimuli to the neural network, such as in a laboratory environment, to determine a particular programming pattern of the memory cells which generates one or more desired operating characteristics of the neural network. This is sometimes referred to as a learning phase for the neural network. It should be noted that one or more methods of facilitating weighting effects in a memory device, such as to affect currents sensed on bit lines of memory devices, is disclosed in U.S. Application Ser. No. 61/625,286, which is commonly assigned.
Operating characteristics of neural networks according to various embodiments of the present disclosure might further be established by configuring the sense circuitry of the memory device. For example, the sense circuitry (e.g., sense circuitry 516, 518 shown in
Particular neuron models have been described comprising strings of memory cells coupled to a single bit line. However, a particular neuron model might be facilitated by memory cell strings coupled to more than one bit line. For example, the current sensed on two or more bit lines might be combined (e.g., summed) to generate a single data output for the particular neuron model comprising the two bit lines. Combining the outputs generated from sensing two bit lines might facilitate improved sensing margin and reduce issues such as current saturation effects during sense operations, for example.
As discussed above,
Similar to combining the output data generated responsive to sensing two or more bit lines to generate a single output data value discussed above, two or more word lines coupled to each string might be biased with input data during a sense operation according to various embodiments of the present disclosure. Remaining word lines coupled to unselected memory cells of the string of memory cells might be biased to a Vpass potential as discussed above. For example, referring to
Although the present example discusses two word lines coupled to two selected memory cells per string being biased during a particular sense operation, the various embodiments are not so limited. For example, one or more word lines coupled to one or more selected memory cells per string of memory cells might be biased during a particular sense operation performed on the neural network, for example. A method of biasing two selected memory cells of the same string of memory cells and facilitating a sense operation on a bit line coupled to the string is disclosed in U.S. Application Ser. No. 61/602,249, which is commonly assigned. As discussed above, remaining memory cells (e.g., unselected memory cells) of each string might be biased to a Vpass potential to operate the unselected memory cells in a pass through mode during each sense operation performed on the neural network, for example.
A programming operation can be performed 602 on memory cells of one or more neural network partitions of the memory device. The memory cells might be programmed as single level memory cells and/or multi level memory cells. The data states of the memory cells might have been previously determined in order to facilitate a particular operating characteristics which the neural network is to exhibit. The partitioning and programming of the memory array is performed prior to operating the memory device in a neural network operating mode according to various embodiments of the present disclosure.
The data programmed in the array of memory cells might be generated by performing an operation to model a system that the neural network is intended to emulate. For example, the desired response characteristics of the system responsive to particular stimuli (e.g., inputs) might be modeled in a laboratory setting. A particular data set can be generated which when programmed into a memory device according to various embodiments of the present disclosure mimics the response of the system modeled in the laboratory. Thus, the operating characteristics of a neural network can be changed by changing the data state of one or more memory cells comprising the neural network, for example.
Subsequent to partitioning 600 and programming 602 the memory array comprising the neural network, the memory device might receive an indication, such as a command from a memory device controller coupled to the memory device, to access the neural network partition of the device. The command might comprise a command to access the neural network along with one or more signals as input values to the neural network. The received input values 604 are used to facilitate a first sense operation on the neuron models of a first level (e.g., LEVEL 1306 shown in
A sense operation is performed 606 on the neural models of the first level of neural models responsive to the received inputs 604. Output signals (e.g., output data) are generated 608 responsive to the sense operation performed 606 on the first level of neural models. The generated output data of the first level of neural models is transferred 610 to a controller, such as controller 530 shown in
A second sense operation 614 of the neural models comprising the second level of neural models can be performed. Output signals (e.g., output data) are generated 616 responsive to the second sense operation performed 614 on the second level of neural models. If it is determined 618 that additional levels of neural models comprise the neural network 620, the generated output signals of the second level of neural models can be output 622 to the controller to facilitate another (e.g., third) sense operation on a third level of neural models of a third level of the neural network, and so on. Thus, according to various embodiments, output data generated during a first sense operation on a first level of a neural network is utilized to determine the inputs to be applied to neuron models of a second level of the neural network during a second sense operation. The output data generated during the second sense operation is utilized to determine the inputs to be applied to neuron models of a third level of neuron models of the neural network, and so on. This process might repeat until no additional levels of neuron models remain to be facilitated in the neural network. Thus, as discussed above, the data state of each memory cell of the neuron models comprising the neural network affect the output data generated during each sense operation thereby defining the operating characteristics of the neural network. Neural networks according to various embodiments can have an associated number representative of the levels of neural models comprising a neural network, such as a number ‘P’ levels of neural models, for example. Thus, a counter might be maintained by the controller where the counter value can be compared to the number P levels of neural models in determining 618 if additional levels of a neural model remain to be sensed, for example. The number of neuron levels P for a neural network might be loaded into the controller during initialization of the memory device, for example.
If it is determined 618 that no additional levels of neural models remain to be facilitated 624, the generated output data of the last level of neural models might be transferred from the neural network 626 as neural network output data. The transferred neural network output data can be transferred to additional circuitry of the memory device, such as data registers or other controlling circuitry configured to store the transferred neural network output data, for example.
The memory device 700 shown in
The memory device 700 of
The number of input data values which can be applied to each neuron model 720-724 comprises in one embodiment the number of strings of memory cells coupled to each bit line and selected for operation in the neural network. Each neuron model comprises a particular number of inputs, such as inputs 202 shown in
The number of neuron models per level might be one through and including N neuron models per level of the neural network, for example. A number of neuron models per level of a neural network according to various embodiments, such as a first level (e.g., LEVEL 1806 shown in
The size of a particular neural network that can be facilitated according to one or more embodiments of the present disclosure might be further described by way of example and reference to
As discussed above with respect to
Following the final sense operation to be performed on a particular neural network, the output data values might be transferred from the word line data register 710 over a bi-directional communications interface 740, such as to other portions of the memory device and/or out of the memory device, such as to a memory access device coupled to the memory device as part of an electronic system, for example. Between each sense operation, the output data values might be transferred to other portions of the memory device.
Application of the various word line potentials might be facilitated by various methods according to one or more embodiments of the present disclosure. The applied potentials might comprise a static application of one or more potentials to the various word lines. For example, one or more word lines coupled to unselected memory cells might be biased with a Vpass potential during a sense operation performed on a selected memory cell of one or more strings of memory cells. The potential applied to selected memory cells can comprise a static potential applied to the selected word lines, for example. According to one or more embodiments, the potential applied to one or more selected memory cells might comprise an increasing (e.g., stepped or ramped) potential or might comprise a decreasing (e.g., stepped or ramped) potential applied during the particular sense operation.
As discussed above, the current sensed on more than one bit line might be combined to generate a single output. For example, an output data value for a particular neuron model might comprise a combination of output data values of two or more bit lines. For example, referring to
Word line drivers, such as those described with respect to
Data corresponding to word line potentials to be applied to particular word lines which is stored in the word line data register 710 might be changed during the execution of a particular memory device operation thereby facilitating applied word line voltages that vary over at least a portion of a duration of the particular memory device operation. For example, these time varying voltages might comprise an increasing or decreasing potential (e.g., stepwise increasing potential and/or stepwise decreasing potential) on one or more word lines during the particular memory device operation. The word line drivers 712 may be further configured to apply one or more time varying voltages on corresponding word lines while concurrently maintaining steady state voltages on one or more different word lines, for example. An example of a steady state potential which might be applied during a read operation might comprise a potential that when applied to memory cells (e.g., unselected memory cells), those memory cells are activated (e.g., exhibit a conductive state) regardless of their data state (e.g., as represented by a programmed threshold voltage) of the memory cells. These potentials facilitate operating one or more unselected memory cells in a pass-through mode.
Each oval region 920-928 shown in
Memory cells comprising oval 930 might be programmed to a data state representative of a threshold level for one or more corresponding neuron models. Each memory cell comprising the oval 930 might be programmed to the same or different data states according to various embodiments of the present disclosure. For example, during a sense operation of a first neuron model, a first memory cell of a string of memory cells comprising the oval 930 and corresponding to the first neuron model might be sensed to determine its data state. The data state determined from the first memory cell of oval 930 might be representative of the overall threshold level of the first neuron model. During a sense operation of a second neuron model, a second memory cell of a string of memory cells comprising the oval 930 and corresponding to the second neuron model might be sensed to determine its data state. The data state determined from the second memory cell of oval 930 might be representative of the overall threshold level of the second neuron model, and so on. The number of data states that memory cells comprising the oval 930 can be programmed to might be the same or might be different than the number of data states that memory cells comprising ovals 920-928 can be programmed to according to one or more embodiments of the present disclosure, for example.
Although shown as single lines to improve readability of the figure, the word lines WLs0-WLs4940 and WLsTS 960 intersecting each oval 920-930 are each representative of multiple word lines (e.g., word line groups) coupled to the one or more strings of memory cells comprising the portions of neuron models comprising each oval. For example, WLs0 shown in
A sense operation according to various embodiments of the present disclosure can be described by way of reference to
During the sense operation, data values indicative of a current or voltage sensed on the bit lines BL0-BL4 and BLS1 are latched in the registers 910 of the sense circuitry 906 coupled to each bit line. The latched data can be transferred 912 from the registers 910 to additional circuitry (not shown) of the memory device. The transferred data can also be transferred to circuitry, such as word line registers and word line drivers (not shown in
Control circuitry 1032 further comprises additional logic circuitry 1036 such as to facilitate a comparison operation of one or more data values generated during sensing, such as those combined together to generate a single data value 1038 for a sense operation performed on a particular neuron model. For example, a single data value 1038 (e.g., summed value) generated as a result of adding currents sensed on bit lines BL0-BL4 might be compared to a data value 1040. The data value 1040 might comprise the threshold level of the particular neuron model. The threshold level of the particular neuron model might be determined by sensing the data state of a memory cell of the oval 1030 which corresponds to the particular neuron model, such as determining a neuron model threshold level from a memory cell of the oval 930 as discussed above with respect to
Control circuitry 1032 might be configured to access an operating characteristic table which defines particular operating characteristics for one or more neuron models and/or the complete neural network. The operating characteristic table might be stored in non-volatile memory of the memory device. The operating characteristic table can be retrieved from the non-volatile memory, such as following a reset operation of the memory device, and be loaded into additional memory (e.g., RAM) (not shown) within the control circuitry 1032. The operating characteristics table might comprise a range of input data values and corresponding range of output data values. The operating characteristic table might facilitate a particular function where the summed value (e.g., summed value of currents sensed in bit lines BL0-BL4) 1038 comprises an input value to the particular function and the output of the particular function comprises the output (e.g., output data value) 1050. The data value (e.g., threshold level) 1040 may also comprise an input variable to the particular function. According to various embodiments, the particular function might comprise one or more linear and/or non-linear functions. For example, the operating characteristics table might facilitate a sigmoid function discussed above. Thus, the control circuitry 1032 can translate an input value (e.g., summed data value 1038) to a corresponding output value (e.g., output 1050) based on the contents of the operating characteristics table.
The operating characteristics table might facilitate a threshold function (e.g., comparator function) where the output data value 1050 comprises a first value when the input (e.g., summed data value 1038) is below a particular threshold level corresponding to the neuron model. When the input to the table is equal to or above the particular threshold level, the output data value comprises a second value, such as facilitating a step function, for example.
The output value 1050 may be transferred to word line registers 1004 as inputs for stimulating another neuron model such as in another level of the neural network, for example. The output value might be transferred 1052 to additional circuitry and/or out of the memory device, such as to a memory access device (e.g., processor) coupled to the memory device, for example.
Row decode and control circuitry 1108 and a column decode and control circuitry 1110 are provided to decode address signals provided to the memory device 1101. Address signals are received and decoded to access memory array 1104. Row decode and control circuitry 1108 further facilitates biasing word lines according to various embodiments of the present disclosure. Row decode and control circuitry 1108 might comprise word line registers, word line drivers and voltage supply circuitry described above, for example.
Memory device 1101 also includes input/output (I/O) control circuitry 1112 to manage input of commands, addresses and data to the memory device 1101 as well as output of data and status information from the memory device 1101. An address register 1114 is coupled between I/O control circuitry 1112 and row decode and control circuitry 1108 and column decode and control circuitry 1110 to latch the address signals prior to decoding. A command register 1124 is coupled between I/O control circuitry 1112 and control logic 1116 to latch incoming commands. Control logic 1116 controls access to the memory array 1104 in response to the commands and generates status information for the external processor 1130. The control logic 1116 is coupled to row decode control circuitry 1108 and column decode control circuitry 1110 to control the row decode control circuitry 1108 and column decode control circuitry 1110 in response to the addresses.
Control logic 1116 can be coupled to a sample and hold circuitry 1118. The sample and hold circuitry 1118 latches data, either incoming or outgoing, in the form of analog data signals. The sample and hold circuitry 1118 can comprise the sensing circuitry and registers comprising the sense circuitry coupled to bit lines as described above with respect to one or more embodiments of the present disclosure, for example. The sample and hold circuitry 1118 could contain capacitors or other analog storage devices for sampling either an incoming data signal representing data to be written to a memory cell or an outgoing data signal indicative of the threshold voltage sensed from a memory cell. The sample and hold circuitry 1118 may further provide for amplification and/or buffering of the sampled signal to provide a stronger data signal to an external device. It is noted that the sample and hold circuitry 1118 could include analog-to-digital conversion functionality and digital-to-analog conversion (DAC) functionality to convert read data from an analog data signal to a digital bit pattern and to convert write data from a digital bit pattern to an analog signal, such as part of sense operations and/or word line potential generation according to various embodiments the present disclosure.
According to various embodiments of the present disclosure, control logic 1116, control circuitry 1112 and/or firmware or other circuitry (e.g., any or all of 1114, 1122, 1124, 1110, and 1108) can individually, in combination, or in combination with other elements, form an internal controller. Control logic 1116 is one example of control circuitry, such as control circuitry 1032 discussed above with respect to
During a write operation, target memory cells of the memory array 1104 can be programmed until voltages indicative of their Vt levels match the levels held in the sample and hold circuitry 1118. This can be accomplished, as one example, using differential sensing devices to compare the held voltage level to a threshold voltage of the target memory cell. Much like traditional memory programming, programming pulses could be applied to a target memory cell to increase its threshold voltage until reaching or exceeding the desired threshold value. In a sense operation, the Vt levels of the target memory cells are passed to the sample and hold circuitry 1118 for transfer to an external processor (not shown in
Threshold voltages of memory cells may be determined in a variety of manners. For example, a word line voltage could be sampled at the point when the target memory cell becomes activated. Alternatively, a boosted voltage could be applied to a first source/drain side of a target memory cell, and the threshold voltage could be taken as a difference between its control gate voltage and the voltage at its other source/drain side. By coupling the voltage to a capacitor, charge would be shared with the capacitor to store the sampled voltage. Note that the sampled voltage need not be equal to the threshold voltage, but merely indicative of that voltage. For example, in the case of applying a boosted voltage to a first source/drain side of the memory cell and a known voltage to its control gate, the voltage developed at the second source/drain side of the memory cell could be taken as the data signal as the developed voltage is indicative of the threshold voltage of the memory cell.
Sample and hold circuitry 1118 may include caching, i.e., multiple storage locations for each data value, such that the memory device 1101 may be reading a next data value while passing a first data value to an external processor, or receiving a next data value while writing a first data value to the memory array 1104. A status register 1122 is coupled between I/O control circuitry 1112 and control logic 1116 to latch the status information for output to the external processor.
Memory device 1101 receives control signals at control logic 1116 over a control link 1132. The control signals may include a chip enable CE#, a command latch enable CLE, an address latch enable ALE, and a write enable WE#. Memory device 1101 may receive commands (in the form of command signals), addresses (in the form of address signals), and data (in the form of data signals) from the processor over a multiplexed input/output (I/O) bus 1134 and output data to the external processor over I/O bus 1134.
In a specific example, commands are received over input/output (I/O) pins [7:0] of I/O bus 1134 at I/O control circuitry 1112 and are written into command register 1124. The addresses are received over input/output (I/O) pins [7:0] of bus 1134 at I/O control circuitry 1112 and are written into address register 1114. The data may be received over input/output (I/O) pins [7:0] for a device capable of receiving eight parallel signals, or input/output (I/O) pins [15:0] for a device capable of receiving sixteen parallel signals, at I/O control circuitry 1112 and are transferred to sample and hold circuitry 1118. Data also may be output over input/output (I/O) pins [7:0] for a device capable of transmitting eight parallel signals or input/output (I/O) pins [15:0] for a device capable of transmitting sixteen parallel signals. The control link 1132 and the I/O bus 1134 might be combined or might be combined in part to form a communications channel between the processor 1130 and the memory device 1101. It will be appreciated by those skilled in the art that additional circuitry and signals can be provided, and that the memory device of
While the memory device of
Methods of facilitating neural network operations in memories, and memories using the methods have been described. In particular, some embodiments facilitate concurrently stimulating one or more memory cells where the one or more memory cells comprise a neuron model. Neuron models can be operated concurrently to be responsive to inputs and generate outputs. Additional embodiments facilitate operating neural networks within a memory device where the neural networks can comprise multiple levels of neuron models. Outputs generated by neuron models of a particular level of the neural network can be supplied as inputs to a different level of neuron models of the neural network thus facilitating a feed-ward type neural network, for example. Methods of establishing operating characteristics of neural networks according to various embodiments have also been described.
Although specific embodiments have been illustrated and described herein, it will be appreciated by those of ordinary skill in the art that any arrangement, which is calculated to achieve the same purpose, may be substituted for the specific embodiments shown. Many adaptations of the disclosure will be apparent to those of ordinary skill in the art. Accordingly, this application is intended to cover any adaptations or variations of the disclosure.
This application is a continuation of U.S. application Ser. No. 16/009,541 filed Jun. 15, 2018, and titled “METHODS OF OPERATING MEMORY,” (allowed), which is a continuation of U.S. application Ser. No. 15/241,496 filed Aug. 19, 2016, and titled “METHODS OF OPERATING MEMORY,” now U.S. Pat. No. 10,068,653 issued on Sep. 4, 2018, which is a continuation of U.S. application Ser. No. 13/774,553 filed Feb. 22, 2013, and titled “NEURAL NETWORK IN A MEMORY DEVICE,” now U.S. Pat. No. 9,430,735 issued on Aug. 30, 2016, which is commonly assigned and incorporated by reference in its entirety. U.S. application Ser. No. 13/774,553 is a non-provisional application of provisional application U.S. Application Ser. No. 61/602,344, filed Feb. 23, 2012, and titled “NEURAL NETWORK IN A MEMORY DEVICE,” and related to provisional application U.S. Ser. No. 61/476,574, filed Apr. 18, 2011, titled “METHODS AND APPARATUS FOR PATTERN MATCHING” which are commonly assigned.
| Number | Date | Country | |
|---|---|---|---|
| 61602344 | Feb 2012 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16009541 | Jun 2018 | US |
| Child | 16515134 | US | |
| Parent | 15241496 | Aug 2016 | US |
| Child | 16009541 | US | |
| Parent | 13774553 | Feb 2013 | US |
| Child | 15241496 | US |
