Computer systems include random-access memories (RAM) for storing data and machine code. RAMs are typically volatile memories, such that the stored information is lost when power is removed. In modern implementations, memories take the form of integrated circuits. Each integrated circuit includes several memory cells. To enable access to stored data and machine code, memories are place in electrical communication with processors. Typically, these electrical communications are implemented as metal traces formed on the substrates on which the memories and the processors are disposed.
Some embodiments relate to a photonic system comprising a plurality of photonic modules including at least first and second photonic modules patterned according to at least one common photomask. Each of the first and second photonic modules comprises first and second boundaries; an optical distribution network; a first optical waveguide optically coupling the optical distribution network to a first neighboring photonic module of the plurality of photonic modules, the first neighboring photonic module being adjacent to the first boundary; and a second optical waveguide optically coupling the optical distribution network to a second neighboring photonic module of the plurality of photonic modules, the second neighboring photonic module being adjacent to the second boundary.
In some embodiments, the first and second boundaries are opposite to one another.
In some embodiments, the first and second optical waveguides are patterned according to the at least one common photomask.
In some embodiments, each of the first and second photonic modules further comprises an out-of-plane optical coupler optically coupled to the optical distribution network.
In some embodiments, the optical distribution network is configured to selectively place the first neighboring photonic module in optical communication with the second neighboring photonic module.
In some embodiments, each of the first and second photonic modules is patterned according to a common set of photomasks, wherein the at least one common photomask is part of the common set of photomasks.
In some embodiments, each of the first and second photonic modules further comprises: third and fourth boundaries, wherein the first and second boundaries are opposite to one another and the third and fourth boundaries are opposite to one another; a third optical waveguide optically coupling the optical distribution network to a third neighboring photonic module of the plurality of photonic modules, the third neighboring photonic module being adjacent to the third boundary; and a fourth optical waveguide optically coupling the optical distribution network to a fourth neighboring photonic module of the plurality of photonic modules, the fourth neighboring photonic module being adjacent to the fourth boundary.
In some embodiments, the optical distribution network is configured to selectively place the first neighboring photonic module in optical communication with the second neighboring photonic module or the third neighboring photonic module.
In some embodiments, the optical distribution network comprises a plurality of optical switches.
In some embodiments, the first and second photonic modules are adjacent to one another such that the second photonic module is the first neighboring photonic module for the first photonic module.
Some embodiments relate to a method for manufacturing a semiconductor wafer comprising: patterning each of at least some of a plurality of photonic modules on the semiconductor wafer using at least one common photomask, wherein patterning each of at least some of the plurality of photonic modules comprises: patterning an optical distribution network; patterning a first optical waveguide optically coupling the optical distribution network to a first neighboring photonic module of the plurality of photonic modules, the first neighboring photonic module being adjacent to a first boundary of the photonic module; and patterning a second optical waveguide optically coupling the optical distribution network to a second neighboring photonic module of the plurality of photonic modules, the second neighboring photonic module being adjacent to a second boundary of the photonic module.
In some embodiments, patterning each of at least some of the plurality of photonic modules comprises patterning the first and second optical waveguides using the at least one common photomask.
In some embodiments, the method further comprises dicing the semiconductor wafer to obtain a photonic substrate comprising: a first photonic module of the plurality of photonic modules; a first neighboring photonic module adjacent to the first boundary of the first photonic module; and a second neighboring photonic module adjacent to the second boundary of the first photonic module.
In some embodiments, the first and second boundaries of the first photonic module are opposite to one another.
In some embodiments, patterning each of at the at least some of the plurality of photonic modules further comprises: patterning a third optical waveguide optically coupling the optical distribution network to a third neighboring photonic module of the plurality of photonic modules, the third neighboring photonic module being adjacent to a third boundary of the photonic module; and patterning a fourth optical waveguide optically coupling the optical distribution network to a fourth neighboring photonic module of the plurality of photonic modules, the fourth neighboring photonic module being adjacent to a fourth boundary of the photonic module. The first and second boundaries are opposite to one another and the third and fourth boundaries are opposite to one another.
In some embodiments, patterning each of at least some of the plurality of photonic modules further comprises: patterning a first photonic module using a first photolithography shot in connection with the at least one common photomask; and patterning a second photonic module using a second photolithography shot, subsequent to the first photolithography shot, in connection with the at least one common photomask.
Some embodiments relate to a computing system comprising a photonic substrate patterned with a plurality of photonic modules including at least first and second photonic modules, each of the first and second photonic modules being patterned according to at least one common photomask, wherein the first photonic module is optically coupled to the second photonic module; a first die in communication with the first photonic module; and a second die in communication with the second photonic module.
In some embodiments, the first die comprises a processor and the second die comprises a memory.
In some embodiments, the computing system further comprises a laser die coupled to the photonic substrate.
In some embodiments, each of the first and second photonic modules comprises: first and second boundaries; an optical distribution network; a first optical waveguide optically coupling the optical distribution network to a first neighboring photonic module of the plurality of photonic modules, the first neighboring photonic module being adjacent to the first boundary; and a second optical waveguide optically coupling the optical distribution network to a second neighboring photonic module of the plurality of photonic modules, the second neighboring photonic module being adjacent to the second boundary.
In some embodiments, the first and second boundaries are opposite to one another.
In some embodiments, the first and second optical waveguides are patterned according to the at least one common photomask.
In some embodiments, each of the first and second photonic module comprises an out-of-plane optical coupler, wherein: the first die is optically coupled to the out-of-plane optical coupler of the first photonic module, and the second die is optically coupled to the out-of-plane optical coupler of the second photonic module.
In some embodiments, the first die is coupled to a first side of the photonic substrate and the second die is coupled to a second side of the photonic substrate opposite the first side.
In some embodiments, the computing system further comprises a third die stacked on top of the first die.
In some embodiments, the first and second photonic modules share a boundary such that the first photonic module is adjacent to the second photonic module.
In some embodiments, the first die is mounted above or below the first photonic module; and the second die is mounted above or below the second photonic module.
In some embodiments, the first die is in electronic communication with the first photonic module, and the second die is in electronic communication with the second photonic module.
Some embodiments relate to a multi-node computing system comprising: a plurality of computing systems including at least first and second computing systems, each of the first and second computing systems comprising: a photonic substrate patterned with a plurality of photonic modules including at least first and second photonic modules, each of the first and second photonic modules being patterned according to at least one common photomask, wherein the first photonic module is optically coupled to the second photonic module; a first die in communication with the first photonic module; a second die in communication with the second photonic module; and a fiber connecting the first and second computing systems to one another.
In some embodiments, each of the first and second computing systems further comprises a fiber coupler, wherein the fiber optically couples respective fiber couplers of the first and second computing systems to one another.
In some embodiments, the first die comprises a processor and the second die comprises a memory.
In some embodiments, each of the first and second computing systems further comprises a laser coupled to the photonic substrate.
In some embodiments, each of the first and second photonic modules of the photonic substrate comprises: first and second boundaries; an optical distribution network; a first optical waveguide optically coupling the optical distribution network to a first neighboring photonic module of the plurality of photonic modules, the first neighboring photonic module being adjacent to the first boundary; and a second optical waveguide optically coupling the optical distribution network to a second neighboring photonic module of the plurality of photonic modules, the second neighboring photonic module being adjacent to the second boundary.
In some embodiments, the first and second optical waveguides are patterned according to the at least one common photomask.
In some embodiments, each of the first and second photonic modules comprises an out-of-plane optical coupler, wherein: the first die is optically coupled to the out-of-plane optical coupler of the first photonic module, and the second die is optically coupled to the out-of-plane optical coupler of the second photonic module.
In some embodiments, the first die is coupled to a first side of the photonic substrate and the second die is coupled to a second side of the photonic substrate opposite the first side.
In some embodiments, the multi-node computing system further comprises a third die stacked on top of the first die.
In some embodiments, the first and second photonic modules share a boundary such that the first photonic module is adjacent to the second photonic module.
In some embodiments, the first die is mounted above or below the first photonic module; and the second die is mounted above or below the second photonic module.
Some embodiments relate to a photonic communication platform comprising: a photonic network comprising a plurality of optical switches formed on a semiconductor substrate; a plurality of dies in communication with the photonic network; an electronic switching network comprising a plurality of transistors co-integrated with the plurality of optical switches, the electronic switching network being configured to: at a first time, program the optical switches to form a first optical communication path coupling together a first subset of the plurality of dies, and at a second time subsequent to the first time, program the optical switches to form a second optical communication path coupling together a second subset of the plurality of dies, the second optical communication path being distinct from the first communication path.
In some embodiments, the plurality of transistors are formed on the semiconductor substrate.
In some embodiments, the semiconductor substrate is a first semiconductor substrate, and wherein the plurality of transistors are formed on a second semiconductor substrate, wherein the first and second semiconductor substrates are 3D-bonded together.
In some embodiments, programming the optical switches to form a first optical communication path comprises: identifying an optical communication path coupling together first subset of the plurality of dies; and programming the optical switches based on the identified optical communication path.
In some embodiments, identifying an optical communication path coupling together first subset of the plurality of dies comprises monitoring a usage of the photonic network.
In some embodiments, the electronic switching network is further configured to determine at least one characteristic of an optical signal at the first optical communication path; identify an encoding scheme based on the at least one characteristic of the optical signal; and cause the photonic network to communicate optically on the first optical communication path based on the encoding scheme.
In some embodiments, the plurality of dies are in electronic communication with the photonic network.
In some embodiments, the electronic switching network is further configured to cause the photonic network to communicate optically on the first optical communication path using wavelength division multiplexing.
Various aspects and embodiments of the application will be described with reference to the following figures. It should be appreciated that the figures are not necessarily drawn to scale. Items appearing in multiple figures are indicated by the same reference number in the figures in which they appear.
The inventors have recognized and appreciated that one of the major bottlenecks limiting the spread of data-intensive computing is the inability to scale memory capacity and bandwidth in modern computers at sufficiently high rates. The inventors have developed photonic communication platforms that enable scaling of memory capacity and bandwidth well beyond what is possible with conventional computers.
Data-intensive computing systems, such as those designed to handle deep learning algorithms, require access to vast amounts of data, thus increasing the requirements for memory capacity. Further, most applications require access to data in real time or quasi real time, thus increasing the requirements for memory bandwidth. Some conventional computing systems utilize graphical processing units (GPUs) to improve memory access efficiency. Some GPUs transfer data from a memory at bandwidths as high as 256 GB/s. While such a memory bandwidth may be sufficient for most graphic-based applications, this is far from being enough for data-intensive applications such as deep neural networks and high-frequency trading.
Deep neural networks rely on large numbers of parameters such as weights and activation parameters. A typical 50-layer network with 26 million weight parameters, for example, can compute up to 16 million activations in a forward pass. If weight and activation parameters are stored using 32-bit floating point values, the total storage requirement is 168 MB. In addition, if data are represented as dense vectors, the memory requirement can increase to several gigabytes. During training, locality of the training dataset is important because of the frequency of access to these large datasets. These amounts of data are far too large to be stored in the internal memory of the GPU, thereby requiring use of several external dynamic random access memories (DRAMs). Each DRAM chip communicates with a processor through electronic communication carriers. For example, in computing systems that utilize silicon interposers, in which the processor and the memory chips are mounted on the same interposer, memory-processor communication is supported by conductive traces formed on the interposer. Use of silicon interposers has become widespread in recent years because the densities of conductive traces that an interposer can provide is much larger than in printed circuit boards (PCBs). However, interposers cannot be scaled indefinitely. Interposers, being manufactured using microfabrication techniques, have limited areas, thus limiting the numbers of memory chips that an interposer can accommodate.
Moreover, the presence of parasitic impedance further limits the bandwidth scalability. Ideally, the impedance of the interposer's conductive traces would be negligible. In reality, unfortunately, the impedance of the traces is significant. The parasitic impedance limits bandwidth scalability in two ways. First, it limits the bandwidth that a trace can support. Second, it increases power consumption. To make things worse, parasitic impedance increases with the length of a trace, meaning that the larger the separation between a memory chip and a processor, the lower the bandwidth. This is why conventional computing systems are typically designed so that memory chips are positioned within a few centimeters of the processor. However, there are only so many memory chips that can be accommodated within this range. The result is that conventional computing systems are limited in both memory bandwidth and memory capacity.
The communication platforms developed by the inventors overcome these limitations using photonics. The physics according to which light propagates inside a waveguide makes optical communications inherently immune to parasitic impedance. The immunity to parasitic impedance leads to a major benefit—it removes the requirement that memory chips be positioned within a certain range of the processor.
Another advantage of the photonic communication platforms developed by the inventors is that they can be easily adapted to different computer architectures. Single-node computer architectures involve one processor chip, which may have more than one processor core per chip, and several memory chips. Multi-node computer architectures involve several processor chips and several memory chips. Some multi-node architectures use a ring topology—each processor is in direct communication with two neighboring processors and communication with other processors passes through the neighboring processors. Other multi-node architectures use a star topology—a central hub is responsible for routing core-to-core communications. Yet another multi-node architectures use a multi-cast topology—each processor is in directed communication to several other processors.
Some aspects of the photonic communication platforms described herein make them easily adaptable to any one of these architectures (and others). Some embodiments provide photonic communication platforms that use “photonic modules.” Each photonic module includes programmable photonic circuits that can be configured based on the needs of a particular computer architecture. Some platforms are arranged according to 1-dimentional schemes, such as in blocks of 3×1 modules, in blocks of 5×1 modules, in blocks of 10×1 modules, 20×1 modules, etc. Some platforms are arranged according to 2-dimentional schemes, such as in blocks of 3×3 modules, in blocks of 5×3 modules, in blocks of 5×5 modules, in blocks of 10×10 modules, etc. More generally, the platforms enable any block of N×M modules, where N≥1 and M≥1, and any topology, such as T-topologies, L-topologies, X-topologies, etc. Each photonic module can serve as a node of a computing system. At each node, there may be one or more digital processor chips, one or more analog accelerators, one or more photonic accelerators, one or more memory chips, one or more networking chips, or other devices.
As described in detail further below, the photonic modules are patterned with optical waveguides and optical distribution networks. The optical distribution network of a photonic module can selectively place the die of that particular photonic module in optical communication with any other die of the computing system. For example, the optical distribution network of the photonic module positioned under processor die 30 may be reconfigured depending on the needs of the processor. At the beginning of a routine, the processor may need to access data stored in a first memory node. This read operation involves configuring the respective optical distribution networks to place the processor in optical communication with the first memory node. Later in the routine, the processor may need to write data into a second memory node. This write operation involves reconfiguring the optical distribution networks to place the processor in optical communication with the second memory node.
The inventors have further appreciated that manufacturing photonic modules in large scales can be costly. The photonic communication platforms described herein are engineered in a way that limits manufacturing costs. These platforms rely on the use of common photomask sets (or at least one common photomask) to fabricate multiple photonic modules. This approach reduces costs in two ways. First, it reduces additional costs that would otherwise be incurred in procuring several different photomask sets. Second, it enables fabrication of photonic modules using standard semiconductor foundries, some of which require that the same photomask set (or at least one photomask) be used across an entire wafer. Designing photonic modules that share at least one photomask enables fabrication of many photonic modules on the same semiconductor wafer while leveraging standard, low-cost step-and-repeat manufacturing processes.
The photonic modules described herein may be manufactured using microfabrication techniques, including for example complementary metal-oxide-semiconductor (CMOS) microfabrication techniques. Accordingly, some embodiments relate to silicon photonics-based optical communication platforms. Some particular microfabrication techniques involve step-and-repeat approaches—whereby stepper machines are used to pattern a semiconductor wafer with multiple copies of a template layout.
Referring first to
Each photomask may define a particular layer of a photonic module. One photomask may be used to define optical waveguides. When the wafer goes through an etch process, only the exposed regions (or only the non-exposed regions) are etched away, while the other regions remain un-etched. This photomask may be patterned to form a network of optical waveguides when the wafer is exposed to light through this photomask.
Some photonic modules involve use of different levels of optical waveguides. In some such embodiments, photomask set 200 may include a dedicated photomask for each waveguide level. Another photomask may be used to define n-doped regions. When the wafer goes through an ion implantation or dopant diffusion process, only the exposed regions (or only the non-exposed regions) receive the doping, while the other regions remain undoped. Another photomask may be used to define p-doped regions using a similar process. Some photonic modules involve use of different doping concentrations. In some such embodiments, photomask set 200 may include a dedicated photomask for each doping concentration. In other embodiments, photomask set 200 may include photomasks used to define deposition of semiconductor materials other than silicon, such as germanium and/or other materials of the periodic table, such as Groups III or V. Another photomask may be used to define metal contacts. Another photomask may be used to define metal traces. Some photonic modules involve use of different levels of metal traces. In some such embodiments, photomask set 200 may include a dedicated photomask for each metal trace level.
In some embodiments, wafer 100 is patterned in a step-and-repeat fashion. When wafer 100 is processed in a stepper machine, the pattern of a photomask is exposed repeatedly across the surface of the wafer, in a grid. This process involves moving the wafer in steps back and forth and left and right under the lens of the stepper, and exposing the photomask at each step. The result is that wafer 100 is patterned with multiple copies of the pattern defined by a photomask. This operation may be repeated for each photomask (or at least some photomasks) of the set.
In some embodiments, this process may be used to pattern wafer 100 with multiple copies of a template photonic module. In the example of
In some embodiments, the entire surface of wafer 100 is patterned using photomask set 200. However, not all embodiments are limited in this respect as some portions of wafer 100 may be patterned using a first photomask set and other portions of wafer 100 may be patterned using a second photomask set. In some embodiments, the first and second photomask sets may share one or more common photomasks, such as a waveguide photomask.
Once patterned, wafer 100 may include multiple photonic substrates. Photonic modules 22 may be separated from the wafer together to form photonic substrates of any desired shape and size. For example, the wafer of
The techniques described in connection with
While the example of
In some embodiments, a photonic module may include multiple layers of photonic waveguides. Similar to how multiple layers of conductive traces increase an electronic circuit's ability to route electric signals, multiple layers of waveguides increase a photonic module's ability to route optical signals. In one example, one layer includes silicon waveguides, and one layer includes silicon nitride waveguides. In another example, multiple layers include silicon waveguides. Additionally or alternatively, multiple layers include silicon nitride waveguides. The choice of material of each waveguide layer may be determined by the wavelength of light that will be routed by the waveguide. For example, silicon and silicon nitride layers may be used for routing infrared light in the telecommunication bands with wavelengths around 1.3 μm or 1.5 μm. In some examples, the multiple layers of waveguides may also include aluminum nitride waveguides that can be used to route visible light down to UV wavelengths or aluminum oxide waveguides that are used to route UV light. Each layer may be arranged in a configuration similar to that illustrated in
Photonic module 22 further includes one or more out-of-plane couplers 105. Waveguide 117 optically couples out-of-plane coupler 105 to optical distribution network 104. Out-of-plane coupler 105 is configured to emit light received from waveguide 117 outside the xy-plane, for example in a direction parallel to the z-axis or at an angle relative to the z-axis. Out-of-plane coupler 105 may be further configured to capture light shining from outside the xy-plane and to convey the captured light to waveguide 117. Out-of-plane coupler 105 enables optical communication between photonic module 22 and a die disposed above the photonic module and/or below the photonic module. Out-of-plane coupler 105 may be implemented using any suitable optical component, including for example optical gratings, lenses, and prisms. In some embodiments, the optical distribution network may be configured so that the same out-of-plane coupler enables optical communication in both directions—from optical distribution network 104 to a die and from the die to optical distribution network 104. In other embodiments, one out-of-plane coupler 105 may enable optical communication in one direction, and another out-of-plane coupler 105 (not shown in
Controller 322 controls the operations of optical distribution network 104. For example, controller 322 controls the state of the switches of optical distribution network 104. Control signals are provided to optical distribution network 104 via electrical connection 324. Alternatively, or additionally, a controller may be formed directly on photonic module 22, and this controller may control the operations of optical distribution network 104. This controller may provide control signals to optical distribution network 22 via conductive traces formed on photonic module 22.
Referring back to
A photonic substrate may include multiple photonic modules connected together to collectively form an optical network.
As discussed above, waveguides of adjacent photonic modules are optically coupled to one another, thereby permitting passage of light from one photonic module to the next. In some embodiments, the waveguides may be physically connected. This arrangement is illustrated in
In other embodiments, there may be a gap between the waveguides. This arrangement is illustrated in
In yet other embodiments, as illustrated in
In some embodiments, photonic modules 22 may be patterned according to a common metal trace photomask. As a result, the photonic modules share the same pattern of metal traces. In some embodiments, photonic modules 22 are patterned according to multiple common photomasks. As a result, multiple levels of metal traces share the same patterns across different photonic modules. Some of the metal traces may be used to deliver power across a photonic substrate. Some of the metal traces may be used to deliver electronic signals across a photonic substrate.
The metal traces are arranged to carry electricity (e.g., signals and/or power) across the boundaries of the photonic modules. This may be achieved by patterning metal traces to be continuous across the boundaries of the photonic modules. In this example, the metal traces of level 1 are continuous across the vertical boundaries, and the metal traces of level 2 are continuous across the horizontal boundaries. Metal traces of different levels may be connected to one another using vias (not shown in
Metal traces may be used to deliver power and/or electric signals across the photonic substrate. In one example, a power source is connected to one particular photonic module. Power produced by the power source may be delivered from that particular photonic module to the other photonic modules using the metal traces. In another example, a controller chip may be bonded (e.g., 3D-bonded) to a particular photonic module. Control signals produced by the controller may be delivered from that particular photonic module to the other photonic modules using the metal traces. The control signals may control the states of the optical distribution networks of the photonic modules.
As discussed above, electronic control circuits may be used to control the operations of the photonic modules. These electronic control circuits may control, for example, how optical distribution networks 104 route optical signals. The electronic control circuits may be co-integrated with the photonic modules in various ways. In some embodiments, a photonic module may be formed on a first substrate and an electronic control circuit may be formed on a second substrate. The two substrates may bonded together to place the electronic control circuit in electrical communication with the optical distribution networks. In other embodiments, however, an electronic control circuit may be fabricated directly on the same substrate as a photonic module. Manufacturing photonic modules and electronic control circuit on the same substrate may reduce costs because, instead of having to rely on two separate fabrication processes and one bonding process, only one manufacturing process may be needed.
Laser die 430 includes one or more lasers. Light produced by the laser(s) can be distributed across the computing system and can serve as reference light to be modulated with data. Laser die 430 can include III-V lasers, such as InP-based lasers. Laser die 430 can be bonded to the photonic substrate, for example using surface mounting techniques. The laser(s) of laser die 430 can be coupled to the semiconductor substrate using out-of-plane couplers. In some embodiments, laser light which is emitted in a direction parallel to the chip surface can be directed towards the out-of-plane couplers using ball lenses.
Recent progress in die-to-wafer bonding of III-V lasers onto a silicon photonics-based wafer shows that the yield of the process may be well below 100%. To circumvent this problem, two or more lasers can be directed towards the same input of the photonic substrate. Only one laser may be used at a time, but if one laser fails, then another laser can be turned on and fed into the photonic substrate. Having multiple lasers per photonic substrate improves the reliability of the platform in case one or more lasers fail.
Chip-based III-V lasers are not the only external light source option for the optical communication platform. Other lasers (e.g., semiconductor-based such as those used in optical telecommunication) can be coupled into the platform using optical fibers or using free-space optics such as lenses. In some embodiments, 1-to-N splitters can be used to enable a single laser to provide light to multiple photonic modules, thus reducing the numbers of laser dies.
Some applications may require multiple lasers and may further require that optical signals emitted by the lasers be mutually coherent (e.g., temporally coherent). In some such embodiments, optical communication platforms of the types described herein can be used to lock the phase of one or more lasers to a single master laser. One or more optical distribution networks 105 may be configured to that the master laser is mixed with one of the slave lasers, and their beat interference pattern is measured. The beat interference pattern is used as an error signal to lock the phases, so that the entire system is coherent.
Lasers emitting at different wavelengths may be used in some embodiments to support wavelength multiplexing schemes. For example, wave-division-multiplexing (WDM) schemes can be used to increase bandwidth utilization per waveguide. Other schemes include multi-mode waveguides, time-division multiplexing and/or polarization diversity. These techniques support multiple independent communication channels using the same optical path.
In some embodiments, appropriate light sources other than lasers such as LEDs or superluminescent diodes may be used in lieu of the aforementioned laser. The choice of the light source is also motivated by the choice of wavelength in the photonic communication fabric. If the fabric is intended to communicate using visible light, then the light sources should be selected to output the appropriate wavelengths of light.
Optical distribution network 104 may be implemented using optical switches. Examples of optical switches include Mach-Zehnder interferometers, optical resonators, multimode interference (MMI) waveguides, arrayed waveguide gratings (AWG), thermos-optic switches, acousto-optic switches, magneto-optic switches, MEMS optical switches, non-linear optical switches, liquid crystal switches, piezoelectric beam steering switches, grating switches, dispersive switches, etc.
Optical distribution networks 104 may be static or dynamic (e.g., reconfigurable based on electrical or optical control signals). A static network may for example receive multiple wavelengths from the same input waveguide and route each wavelength to a different output waveguide. Another static network may receive two orthogonal polarizations from the same input waveguide and route each polarization to a different output waveguide. Another static network may receive multiple modes from the same input multimode waveguide and route each mode to a different output waveguide.
Dynamic optical distribution networks may be reconfigured depending upon the needs of a computing system.
A fully connected routing topology, however, may not be necessary or feasible in some embodiments. To reduce the complexity of the optical distribution network, routing may be restricted to fewer options, as shown in the example of
The inventors have appreciated that interfacing dies (e.g., memories, processors, etc.) with the photonic substrates described herein raises compatibility issues. Ideally, dies are predefined with control circuits for controlling the operations of the optical distribution networks and pins that align perfectly with the pins of the photonic substrate. In this way, the die and the photonic substrate are inherently predisposed to communicate with each other once they are bonded together. This approach, however, may be impractical because the die and the photonic substrate are often manufactured by different entities. For example, one entity in the United States may manufacture the photonic substrate and another entity in Japan may manufacture a die. This approach puts the burden on the entity that manufactures the die to include control circuits and pins that are compatible with the photonic substrate, which can substantially increase costs for the die manufacturer. Recognizing this problem, the inventors have developed electronic switching networks that serve as interfaces between photonic substrates and dies. These electronic switching networks are arranged so that dies need not be redesigned to be compatible with a photonic substrate, thus saving costs to die manufacturers. In essence, an electronic switching network defines the switching and control protocols for a photonic communication platform.
Some electronic switching networks are co-fabricated with the photonic modules. In some embodiments, an electronic switching network is formed on the same substrate on which the photonic modules are formed. Referring back to
The electronic switching network may program the optical distribution networks based on data stored in the memories 742. For example, the memories may store instructions that, when executed, cause the optical distribution networks to perform a predefined switching sequence. Alternatively, the memories may store instructions that, when executed, cause the optical distribution networks to dynamically optimize the optical links based on the needs of a particular application. In some embodiments, the electronic switching network may monitor usage of the optical links, and may determine which optical links are being used and which optical links are available, or which optical links can provide a bandwidth in excess of a threshold bandwidth. The electronic switching network may use this information to decide how to allocate optical links to particular data streams. This may be particularly useful in photonic communication platforms connecting, for example, hundreds of memory dies and processors. Such photonic communication platforms can rely on the electronic switching network to determine the optimal optical path allowing a particular processor die to access information from a particular memory die. Machine learning algorithms may be used in some embodiments to identify optimal optical links.
In some embodiments, an electronic switching network is configured to perform the following steps. At a first time, the electronic switching network programs the optical switches of a photonic substrate to form a first optical communication path coupling together a first subset of the plurality of dies (e.g., coupling a first die to a second die). Programming the optical switches at the first time may involve changing the states of some of the switches (e.g., a first subset of the switches) without changing the states of other switches, or may involve changing the states of all the switches. At a second time subsequent to the first time, the electronic switching network programs the optical switches to form a second optical communication path coupling together a second subset of the plurality of dies (e.g., a third die to the first die or a third die to a fourth die). The electronic switching network may continue to reprogram the optical switches as needed. Programming the optical switches at the second time may involve changing the states of some of the switches (e.g., the first subset of the switches or a second subset of the switches) without changing the states of other switches, or may involve changing the states of all the switches. Changing the state of a switch may involve any of the following operations: changing an outputs of the switch from an off-state to an on-state, changing an output of the switch from an off-state to a partially on-state (e.g., a state in which the switch delivers a fraction of the input power to the output and a fraction of the input power to one or more other outputs), changing an output of the switch from an on-state to a partially on-state, changing an output of the switch from an on-state to an off-state, selecting one or more inputs of the switch that were previously unselected, and/or unselecting one or more inputs of the switch that were previously selected.
In the example of
A typical connection between nodes may begin with a request message on the electronic switching network. The request may be arbitrated through a mesh of electronic switches to establish a connection. At each network hop, the optical link is made once the electrical request wins arbitration. Once the connection is fully established between the source and destination, an acknowledgement is returned to the requester on the electronic network and data transmission proceeds over the optical link.
Controllers 740 may use debugging units 744 to perform testing, diagnosis and fault isolation of metal traces, contacts, pins, pads, etc. In some embodiments, the debugging units are implemented using the Joint Test Action Group (JTAG) industry standard. In some embodiments, manufacturing defects could result in non-functional communications channels. Redundant connections (optical and/or electrical) can be included and mapped in as replacements so that platforms with small numbers of defects need not be discarded.
The ability to dynamically reconfigure a photonic communication platform in a dynamic fashion allows the electronic switching network to operate at relatively low bandwidths, thus leading to relatively low power consumption. Being low power is particularly beneficial when the amount of data carried over the photonic modules between switching events is relatively large. Being low power, in some embodiments, opens the opportunity for “always-on” operations that can monitor training or reconfiguration sequences.
In some embodiments, the power and fidelity of an optical signal that propagates across multiple photonic modules may depend upon the number of waveguide crossings that the signal crosses. Thus, the farther an optical signal has to travel the larger the power consumption and the lower the fidelity. Recognizing this issue, the inventors have developed configurations that can adaptively select encoding schemes with larger or smaller alphabets (more or less bits per symbol) depending upon the number of photonic module boundaries that a signal is supposed to cross or an expected optical loss, or any other characteristic of an optical path. For example, when only a few crosses or small amounts of losses are expected, an encoding scheme with a larger alphabet can be used. Examples of such encoding schemes include quadrature amplitude modulation (QAM) that use a constellation of 16 points (16-QAM), 64 points (64-QAM), or 256 points (256-QAM), and pulse-amplitude modulation (PAM) with many levels, e.g. PAM-16 or PAM-32. Vice versa, when many crosses or large amounts of losses are expected, an encoding scheme with a smaller alphabet can be used. Examples of such encoding schemes include binary phase shift keying (BPSK) or PAM with few levels, e.g. PAM-2 or PAM-4, In some embodiments, the bandwidth of an optical path may be set depending upon changing communication patterns, or depending upon the number of photonic module boundaries that a signal is supposed to cross.
Clocking of the electronic switching network can be done using a single distributed clock. The clock phase may be tuned in the photonics domain to achieve extremely low jitter at each receiver. This avoids the need for an embedded clock and thus the data may not need to be encoded. Non-encoded data allows for increased bandwidth, reduced latency, and reduced power.
Computing systems may be formed that leverage the photonic communication platforms described herein. Unlike electronic distribution networks, these optical communication platforms can provide copies of the same message to multiple locations simultaneously without incurring problems arising due to parasitic impedance. This characteristic allows the optical distribution networks to form multicast and/or broadcast communications schemes. The optical distribution networks can be dynamically reconfigured to route messages to a single node or multiple nodes. Leveraging the broadcast and/or multicast capability, some embodiments enable performance of MapReduce operations directly using an optical communication platform.
Computing systems of these types may be used in a variety of applications, including for example in high-performance computing, neural, machine learning and deep learning networks, graphics rendering, large-scale visualization, gaming, high-frequency trading and video streaming, among others.
The bottom inset of
The top inset of
The computing systems described in connection with
The inventors have further appreciated that optical modules of the types described herein may serve as transceivers enabling communication from a computing system to another computing system or from a computer system to a host. This transceiver may be optical or electronic. Example optical interfaces include board-to-board communication via optical fibers or remote rack-to-rack communication utilizing higher level protocols such as Ethernet or Infiniband. Electronic interfaces to a host system include SERDES-based standards such as PCI Express. An external I/O module may manage communications between the host and a local system. This includes direct memory access offload capabilities for high speed data movement between remote and local memories. The external I/O module may further provide local interfaces for management, calibration, boot, and reliability and serviceability (RAS).
Optical communication platforms of the types described herein can provide layered network stacks. One example stack is configured as follows. A physical layer includes optical interconnects including optical-to-electrical converters, electrical-to-optical converters, and optical distribution networks. A datalink layer includes an electronic switching network allowing connections between modules. In some embodiments, communication between nodes of a network may begin with a request message on the electronic switching network. The datalink layer processes this request and arbitrates through a mesh of electrical switches to establish a connection. At each photonic module boundary, an optical link is formed if the request wins the arbitration. Once the connection is fully established between the source and the destination, an acknowledgement may be returned to the requester on the electronic switching network and data transmission proceeds over the photonic modules. A transport layer involves packetization, data integrity, and buffer allocation. The transport layer uses packetization to implement upper layer protocols on the datalink layer. Flow control may be implemented with buffer credits. If reliability dictates additional protection from errors at the physical layer, data integrity including cyclic redundancy checks (CRC) (e.g., resend) and/or forward error correction (FEC) schemes can be used.
Having thus described several aspects and embodiments of the technology of this application, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those of ordinary skill in the art. Such alterations, modifications, and improvements are intended to be within the spirit and scope of the technology described in the application. It is, therefore, to be understood that the foregoing embodiments are presented by way of example only and that, within the scope of the appended claims and equivalents thereto, inventive embodiments may be practiced otherwise than as specifically described. In addition, any combination of two or more features, systems, articles, materials, and/or methods described herein, if such features, systems, articles, materials, and/or methods are not mutually inconsistent, is included within the scope of the present disclosure.
Also, as described, some aspects may be embodied as one or more methods. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.
All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and/or ordinary meanings of the defined terms.
The indefinite articles “a” and “an,” as used herein in the specification and in the claims, unless clearly indicated to the contrary, should be understood to mean “at least one.”
The phrase “and/or,” as used herein in the specification and in the claims, should be understood to mean “either or both” of the elements so conjoined, i.e., elements that are conjunctively present in some cases and disjunctively present in other cases.
As used herein in the specification and in the claims, the phrase “at least one,” in reference to a list of one or more elements, should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each and every element specifically listed within the list of elements and not excluding any combinations of elements in the list of elements. This definition also allows that elements may optionally be present other than the elements specifically identified within the list of elements to which the phrase “at least one” refers, whether related or unrelated to those elements specifically identified.
The terms “approximately” and “about” may be used to mean within ±20% of a target value in some embodiments, within ±10% of a target value in some embodiments, within ±5% of a target value in some embodiments, and yet within ±2% of a target value in some embodiments. The terms “approximately” and “about” may include the target value.
The present application is a Continuation claiming the benefit under 35 U.S.C. § 120 of U.S. patent application Ser. No. 17/313,415, filed May 6, 2021, entitled “PHOTONIC COMMUNICATIONS PLATFORM,” which is a Divisional claiming the benefit under 35 U.S.C. § 120 of U.S. patent application Ser. No. 16/810,573, filed Mar. 5, 2020, entitled “PHOTONIC COMMUNICATION PLATFORM,” which claims the benefit under 35 U.S.C. § 119(e) of U.S. Provisional Patent Application Ser. No. 62/961,448, entitled “PHOTONICS COMMUNICATIONS PLATFORM WITH SINGLE LITHOGRAPHIC MASK SET,” filed on Jan. 15, 2020; U.S. Provisional Patent Application Ser. No. 62/923,889, entitled “PHOTONICS COMMUNICATIONS PLATFORM WITH SINGLE LITHOGRAPHIC MASK SET,” filed on Oct. 21, 2019; and U.S. Provisional Patent Application Ser. No. 62/814,444, entitled “PHOTONICS COMMUNICATIONS PLATFORM WITH LITHOGRAPHIC MASK SET,” filed on Mar. 6, 2019, each of which is hereby incorporated herein by reference in its entirety.
Number | Name | Date | Kind |
---|---|---|---|
5416861 | Koh et al. | May 1995 | A |
5771323 | Trott | Jun 1998 | A |
5930429 | Trott | Jul 1999 | A |
6202165 | Pine | Mar 2001 | B1 |
6477285 | Shanley | Nov 2002 | B1 |
7251386 | Dickinson et al. | Jul 2007 | B1 |
7310459 | Rahman | Dec 2007 | B1 |
7894699 | Beausoleil | Feb 2011 | B2 |
8450186 | Rong et al. | May 2013 | B2 |
9671572 | Decker et al. | Jun 2017 | B2 |
9922887 | Vermeulen et al. | Mar 2018 | B2 |
10847467 | Collins et al. | Nov 2020 | B2 |
11036002 | Harris | Jun 2021 | B2 |
11754783 | Harris et al. | Sep 2023 | B2 |
11860413 | Harris et al. | Jan 2024 | B2 |
20020028045 | Yoshimura et al. | Mar 2002 | A1 |
20040258408 | Ramaswami et al. | Dec 2004 | A1 |
20050224946 | Dutta | Oct 2005 | A1 |
20070280585 | Warashina et al. | Dec 2007 | A1 |
20080044128 | Kish, Jr. et al. | Feb 2008 | A1 |
20080089640 | Beausoleil | Apr 2008 | A1 |
20090103345 | Mclaren et al. | Apr 2009 | A1 |
20090103855 | Binkert et al. | Apr 2009 | A1 |
20100054671 | Ban et al. | Mar 2010 | A1 |
20110052120 | Tan et al. | Mar 2011 | A1 |
20110073989 | Rong et al. | Mar 2011 | A1 |
20110128790 | Sarin et al. | Jun 2011 | A1 |
20120149148 | Dallesasse et al. | Jun 2012 | A1 |
20120177381 | Dobbelaere et al. | Jul 2012 | A1 |
20120203695 | Morgan et al. | Aug 2012 | A1 |
20130051725 | Shinoda et al. | Feb 2013 | A1 |
20130071121 | Sharapov et al. | Mar 2013 | A1 |
20130156366 | Raj et al. | Jun 2013 | A1 |
20130209112 | Witzens | Aug 2013 | A1 |
20130243429 | Whelihan et al. | Sep 2013 | A1 |
20130292840 | Shoemaker et al. | Nov 2013 | A1 |
20130308893 | Zuffada et al. | Nov 2013 | A1 |
20140040698 | Loh et al. | Feb 2014 | A1 |
20140043050 | Stone et al. | Feb 2014 | A1 |
20140264400 | Lipson | Sep 2014 | A1 |
20140268980 | Kim et al. | Sep 2014 | A1 |
20140294342 | Offrein et al. | Oct 2014 | A1 |
20140300008 | Duan et al. | Oct 2014 | A1 |
20140363124 | Pelley et al. | Dec 2014 | A1 |
20140363172 | Pelley et al. | Dec 2014 | A1 |
20150381273 | Gloeckner et al. | Dec 2015 | A1 |
20160085038 | Decker et al. | Mar 2016 | A1 |
20160181322 | Mazed et al. | Jun 2016 | A1 |
20160191188 | Butler | Jun 2016 | A1 |
20160216445 | Thacker et al. | Jul 2016 | A1 |
20160252688 | Barwicz et al. | Sep 2016 | A1 |
20170108655 | Zarbock et al. | Apr 2017 | A1 |
20170160474 | Mahmoodian et al. | Jun 2017 | A1 |
20170194309 | Evans et al. | Jul 2017 | A1 |
20180045885 | Canali et al. | Feb 2018 | A1 |
20190162901 | Yu et al. | May 2019 | A1 |
20190189603 | Wang et al. | Jun 2019 | A1 |
20190310433 | Yoo et al. | Oct 2019 | A1 |
20190335252 | Ryan | Oct 2019 | A1 |
20190363797 | Peterson et al. | Nov 2019 | A1 |
20200006304 | Chang et al. | Jan 2020 | A1 |
20200111720 | Wan et al. | Apr 2020 | A1 |
20200284981 | Harris et al. | Sep 2020 | A1 |
20210096311 | Yu et al. | Apr 2021 | A1 |
20210118853 | Harris et al. | Apr 2021 | A1 |
20210215897 | Epitaux et al. | Jul 2021 | A1 |
20210242124 | Kannan et al. | Aug 2021 | A1 |
20210278590 | Harris et al. | Sep 2021 | A1 |
20210375829 | Or-Bach et al. | Dec 2021 | A1 |
20220109075 | Byrd et al. | Apr 2022 | A1 |
20220148627 | Meade et al. | May 2022 | A1 |
20230085268 | Harris et al. | Mar 2023 | A1 |
20230114847 | Harris et al. | Apr 2023 | A1 |
20230308188 | Dorta-Quinones et al. | Sep 2023 | A1 |
20230314711 | Eslampour et al. | Oct 2023 | A1 |
20230314742 | Dorta-Quinones et al. | Oct 2023 | A1 |
20230388024 | Tymchenko et al. | Nov 2023 | A1 |
20230400632 | Harris et al. | Dec 2023 | A1 |
20230408764 | Harris et al. | Dec 2023 | A1 |
Number | Date | Country |
---|---|---|
104092500 | Oct 2014 | CN |
106817323 | Jun 2017 | CN |
0547262 | Jun 1993 | EP |
2 212 887 | Apr 2013 | EP |
3 159 721 | Apr 2017 | EP |
H07-98463 | Apr 1995 | JP |
2005-502127 | Jan 2005 | JP |
2011-501238 | Jan 2011 | JP |
2011-503760 | Jan 2011 | JP |
2015-062027 | Apr 2015 | JP |
2018-506072 | Mar 2018 | JP |
2018-093007 | Jun 2018 | JP |
2018-195723 | Dec 2018 | JP |
10-2010-0095515 | Aug 2010 | KR |
508834 | Nov 2002 | TW |
WO 2011143548 | Nov 2011 | WO |
WO 2016008771 | Jan 2016 | WO |
WO 2016124762 | Aug 2016 | WO |
2019132970 | Jul 2019 | WO |
Entry |
---|
International Search Report and Written Opinion mailed Nov. 29, 2023, in connection with International Application No. PCT/US23/65007. |
Singapore Search Report and Written Opinion dated Dec. 1, 2023, in connection with Singapore Application No. 11202108868T. |
Orcutt et al., Open foundry platform for high-performance electronic-photonic integration. Optics Express.May 21, 2012;20(11). 11 pages. |
International Search Report and Written Opinion mailed Jan. 18, 2023, in connection with International Application No. PCT/US2022/046379. |
International Preliminary Report on Patentability mailed Aug. 18, 2023, in connection with International Application No. PCT/US2021/016129. |
International Search Report and Written Opinion mailed Jan. 26, 2023, in connection with International Application No. PCT/US2022/043209. |
Singapore Search Report and Written Opinion dated May 13, 2022, in connection with Singapore Application No. 11202011400P. |
Extended European Search Report dated Nov. 4, 2022, in connection with European Application No. 20766814.6. |
Invitation to Pay Additional Fees mailed Jun. 19, 2019, in connection with International Application No. PCT/US2019/029803. |
Invitation to Pay Additional Fees mailed Nov. 7, 2022, in connection with International Application No. PCT/US2022/043209. |
Extended European Search Report dated Feb. 11, 2022, in connection with European Application No. 19803311.0. |
International Search Report and Written Opinion mailed Jul. 6, 2020, in connection with International Application No. PCT/US2020/021209. |
International Search Report and Written Opinion mailed Sep. 3, 2019, in connection with International Application No. PCT/US2019/029803. |
Invitation to Pay Additional Fees mailed May 1, 2020, for International Application No. PCT/US2020/021209. |
International Preliminary Report on Patentability mailed Nov. 26, 2020, in connection with International Application No. PCT/US2019/029803. |
International Preliminary Report on Patentability mailed Sep. 16, 2021, in connection with International Application No. PCT/US2020/021209. |
International Search Report and Written Opinion mailed Apr. 14, 2021, in connection with International Application No. PCT/US2021/016129. |
Invitation to Pay Additional Fees dated Aug. 25, 2023, in connection with International Application No. PCT/US2023/065007. |
Beausoleil et al., Nanoelectronic and nanophotonic interconnect. Proceedings of the IEEE. Feb. 2008;96(2):230-47. doi:10.1109/JPROC.2007.911057. |
Young et al., Optical technology for energy efficient I/O in high performance computing. IEEE Communications Magazine. Oct. 2010;48:184-91. |
Fujikata et al., LSI on-chip optical interconnection with Si nano-photonics. Extended Abstracts of the 2007 International Conference on Solid State Devices and Materials. 2007. pp 276-277. |
Grigalunas, Tell Me—What Is Wafer Dicing? Eastern States Components, LLC. Sep. 26, 2017. 2 pages. URL:https://www.escomponents.com/blog/2017/9/26/tell-me-what-is-wafer-dicing [retrieved Jul. 13, 2020]. |
Wada, Electronics and Photonics Convergence on Si CMOS Platform. Proc. of SPIE. 2004;5357:16-24. |
Udipi et al., Combining memory and a controller with photonics through 3D-stacking to enable scalable and energy-efficient systems. 38th Annual International Symposium on Computer Architecture (ISCA). Jun. 8, 2011. pp. 425-436. |
U.S. Appl. No. 18/455,235, filed Aug. 24, 2023, Harris et al. |
U.S. Appl. No. 18/455,395, filed Aug. 24, 2023, Harris et al. |
U.S. Appl. No. 18/526,714, filed Dec. 1, 2023, Harris et al. |
U.S. Appl. No. 18/526,652, filed Dec. 1, 2023, Harris et al. |
PCT/US23/65007, Nov. 29, 2023, International Search Report and Written Opinion. |
11202108868T, Dec. 1, 2023, Singapore Written Opinion. |
Bell, Shane et al., TILE64—Processor: A 64-Core SoC with Mesh Interconnect, 2008 International Solid-State Circuits Conference—Digest of Technical Papers, 200802. pp. 1-3. DOI:10.1109/ISSCC.2008.4523070. |
International Search Report and Written Opinion dated Feb. 28, 2024, in connection with International Application No. PCT/US23/80883. |
Extended European Search Report dated Feb. 19, 2024, in connection with European Application No. 21750009.9. |
PCT/US23/80883, Feb. 28, 2024, International Search Report and Written Opinion. |
EP 21750009.9, Feb. 19, 2024, Extended European Search Report. |
Number | Date | Country | |
---|---|---|---|
20230358957 A1 | Nov 2023 | US |
Number | Date | Country | |
---|---|---|---|
62961448 | Jan 2020 | US | |
62923889 | Oct 2019 | US | |
62814444 | Mar 2019 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 16810573 | Mar 2020 | US |
Child | 17313415 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 17313415 | May 2021 | US |
Child | 18356680 | US |