Patent Application
20220413312
This application relates generally to optical systems for communication in computational systems, and more specifically to techniques and devices for computing with shuffle-exchange systems using optical systems to perform the shuffle.
Computational systems are increasingly limited by the wires used to interconnect processing elements. The signal delay for the longer wires dominates the gate delay. These wires consume more space on the system as the computational systems grow in complexity. And as the feature size is made smaller, the power dissipated by a long-distance wire dominates the power required for a gate.
The on-chip long-wire interconnect requires buffered segments to overcome a delay which otherwise grows quadratically with wire length. With periodic buffering, the delay and power required for a wired interconnect still grow linearly with interconnect length.
The industry is searching for an interconnect technology which has high density, has minimal delay and has a power consumption level that grows more slowly than the interconnection length.
Optical interconnections for VLSI have been explored in both waveguide form and free space form. Efforts to utilize optical waveguides provide high speed optical interconnects, but these cannot easily provide a number of channels which scales with chip area, since the channels usually exit the chip edge.
Free space optical interconnects have been proposed, which route data out the surface of the chip. For regular interconnection patterns, these can have a very high density, approaching the maximum given by the space-bandwidth product of the system. The versions of these that utilize space-variant interconnection patterns do not have this high density, but they allow a choice of interconnects not available with the regular interconnection patterns.
Optical shuffle-exchange networks take advantage of the long-range connection capabilities of optics by performing the shuffle in the optical system. Combined with arbitrarily configurable exchange subnetworks, a multistage version of these systems would be able to perform space variant interconnects. However large, configurable exchange subnetworks have significant complexity in themselves, because an exchange network with full degrees of freedom is combinatorially complex. Additionally, optical shuffle networks, to date, have not been developed with density comparable to the available space-bandwidth product of the optical system, due to underfilled subarrays in the shuffle, and systems that can't be tiled into larger networks.
Consequently, the industry is searching for a high-density optical interconnect technology which has minimal delay, low power consumption and which can maintain high density interconnects while scaling with increasingly complex computational systems.
The present invention provides an optical network that allows effective routing of signals between collected arrays of ports and dispersed arrays of ports on electronic computing chips (reversibly, also between the dispersed arrays and the collected arrays).
Apparatus for optical communication includes an originating chip segment comprising an array of originating tiles, which each comprise arrays of transmitters configured to transmit beamlets perpendicular to the tiles and a terminating chip segment comprising an array of terminating tiles, the terminating tiles each comprising arrays of receivers. Each transmitter corresponds to a particular receiver.
An originating lenslet array having two parallel layers of lenslets is parallel and adjacent the array of originating tiles. Each lenslet pair within the lenslet array pair corresponds to a tile. The lenslet array pair collimates beams from transmitters such that the location of a transmitter within its tile determines the direction of its resulting collimated beam.
Similarly, a terminating lenslet array pair focuses collimated beams to receivers such that the direction of a collimated beam determines which receiver it is focused on. Thus, beamlets transmitted from the transmitters pass through the originating lenslet array and the terminating lenslet array such that the beamlet from each transistor arrives at its corresponding receiver.
In a useful embodiment, the originating chip segment comprises collected tiles, and the terminating chip segment comprises dispersed tiles (designated itels).
Generally, the beamlet from each transmitter within an originating tile is directed to a different terminating tile. For example, if each transmitter and each receiver has a pixel number within its tile and a tile number within its array, and then the pixel number and tile number of each transmitter is a transpose between the pixel number and the tile number of its corresponding receiver.
As a feature, the originating transmitters are densely spaced such that the distance between adjacent transmitters on a tile is on the order of an f-number of the originating lenslet array pair times a wavelength of light being transmitted by the transmitters. Similarly, the distance between adjacent originating tiles is much smaller than the width of a tile.
In some systems, external illumination is provided and the transmitters are modulators. It is useful to have the modulators illuminated by circularly polarized light.
Systems also often operate in the reverse direction, meaning the transmitters are transmitter-receivers, the receivers are receiver-transmitters, and a beam from a receiver-transmitter is directed to its corresponding transmitter-receiver. Both can be modulators.
Optical communications systems may also be used to shuffle data. For example the terminating chip converts beams received at a first set of receiver-transmitters to electric signals, routes the electric signals to a second set of receiver-transmitters and converts the electric signals back to optical signals for transmission by the second set of receiver-transmitters to their corresponding transmitter-receivers.
An embodiment of the invention includes a free-space optical train disposed between the collected array of tiles and the dispersed array of tiles and an array of optoelectronic transducers (pixels) on each tile for converting back and forth between optical and electrical signals. A tile is a collection of transducers that sends and/or receives light through a common lenslet or lenslet pair. The free-space optical train includes a lenslet array to collimate the optical signals, a lenslet array to focus the optical signals and a system for routing the optical signals from the collimating lenslets to the focusing lenslets. The system for routing the optical signals can include a Fourier transform lens to map the array of angles from each collimating lens in the collimating lenslet array to the array of positions making up the focusing lenslet array. The free-space optical train may perform a shuffle rearrangement of the signals. The system for routing the optical signals can be incorporated into the focusing and collimating lenslet arrays. The optoelectronic transducers for converting back and forth between optical and electrical signals may be emitters that can also be used as detectors. The optoelectronic transducers may be optical modulators that can also be used as detectors. The free-space optical train may include beam splitters, for example, to accept light from an external source and route it to the modulators acting as optoelectronic transducers. The electronic signals on tiles of the dispersed array may be rearranged by electrical connections (exchanged) and converted back to an optical signal transmitted to the collected array. Transmission of the optical signals from the dispersed array back the collected array may traverse the same optical system with the focusing lenslet array acting as a collimating array and the collimating array acting as a focusing array in the back transmission of the signals. Herein the term transmitter includes modulators.
The invention includes dynamic (rearrangement) embodiments and static embodiments. In dynamic embodiments, the signals on the dispersed array are a rearranged with time dependent patterns. In static embodiments, the path of the signals on the dispersed array remains the same during a computation.
In some embodiments of the invention, the rearrangement of signals on the dispersed array tiles is achieved by routing the signals with a nearest-neighbor-interconnected store and forward connection. In other embodiments, the rearrangement of signals may be done with long and short interconnections between elements on the dispersed array tiles.
In some embodiments of the invention the dispersed array contains memory registers used for storing data for the computational system.
In some embodiments of the invention, the collected array contains blocks of computational elements arranged in systolic arrays of multiply-accumulators for vector-matrix multiply and matrix-matrix multiply operations.
In some embodiments, a block of data is transmitted from a source collected array tile to a destination collected array tile by transmitting it to a block of dispersed array tiles and providing that each of the dispersed array tiles in the block move the signal with the same pattern as each other, before transmitting the signals back to the collected array tiles. The block of data is moved a long range across the collected tiles by moving each signal within the block a small range on the dispersed tiles.
It is desirable to build the optical transducers as telecentric, high-numerical aperture modulators or emitters, since this allows the highest density of signal channels. It is also desirable to provide a transversely compact optical system constituting the lenslet for each tile, so that the lenslet array to collimate the optical signals, and the lenslet array to focus the optical signals remain compact, and consequently allow that the dead space between the tiles remains small. It is also desirable to provide a transversely compact optical system for routing the optical signals between the collimating lenslets and the focusing lenslets, so that multiple free-space optical trains addressing their own arrays of tiles may be laid side to side and connected electronically with short electrical interconnects.
An application where the computational elements arranged in systolic arrays of multiply-accumulators might involve the use of the matrix-matrix multiplier to compute activity in an artificial neural network.
A system for communicating between VM multipliers (computational elements) with an intermediate stage may include features where the intermediate stage has high memory capacity or the intermediate stage has routing. The VM may be matrix-matrix multipliers or systolic arrays.
An optical system is taught for performing an N-way shuffle (transpose), performing the same permutation (possibly with copying) on each tile of the shuffled plane, and performing the N-way shuffle back to the original plane, for moving/copying data in large blocks. The shuffled plane can contain high memory density for storing intermediate results
The following description sets forth embodiments of an optical shuffle computation network according to the invention. Embodiments of the invention can be applied to computational environments such as artificial neural network systems, array computation, memory access and to networked computational systems.
The general functionality of the optical shuffle computation network is to accept signals from an array of collected tiles, transmit signals from each over a long range to a set of dispersed tiles, where the signals may access memory and where the signals may be exchanged locally and transmitted back to the original or to an exchanged collected tile.
Originating transducer array has four tiles 410, 420, 430, and 440 with nine pixels or transducers (receivers, transmitters, modulators, etc.) each (in the one dimension shown). The second pixel in the first tile 410 is pixel 412. In general, pixels are designated something like Ptile,pixel so pixel 412 is OP1,2 in that notation. To avoid confusion in this description, a different reference number is used for each pixel discussed. Pixel 437 is seventh pixel in the third tile, 430. It could be designated OP3,7.
Terminating transducer array 600 has nine tiles 610, 620, 630, 640, 650, 660, 670, 680, and 690 with 4 pixels each. Pixel 621 (TP2,1) is the first pixel in the second tile and pixel 673 (TP7,3) is the third pixel in the seventh tile 670.
This embodiment may be bidirectional. In this discussion, transmission originates from collected array 400, so transducers 401-440 on the collected side are sometimes designated transmitter-receivers. The signals originating from array 400 terminate at dispersed array 600, and so transducers 601-690 on the dispersed side are sometimes designated receiver-transmitters.
Quarter-wave plates 40 and 60 rotate polarization of light from light sources 800, 100, so that after reflective modulation the light passes straight through polarizing beam splitters 31 and 30 from array 400 to array 600 and (in a two-way system) from array 600 to array 400. Optoelectronic transducers on array 400 and array 600 may operate as reflective modulators in signal transmitting mode, or as detectors in signal receiving mode.
Two-layer lenslets in arrays 500 and 700 operate as Fourier transform pairs for the tiles in arrays 400 and 600. Each tile faces a pair of lenslets (e.g. tile 410 faces lenslets 510 and 511). The cone of light from the transducer 412 is turned by the first lenslet 510 and the second lenslet collimates the light, directed at an angle corresponding to the position of the transducer. The lenslets 510 and 511 have a common focal length and are separated by a distance equal to the focal length. The lens pair 44 and 64 also have a common focal length and are separated by a distance equal to the focal length, making an additional Fourier transform pair. The transmitters may be densely packed, meaning that the distance between adjacent transmitters on a tile is on the order of an f-number of the originating lenslet array pair times a wavelength of light being transmitted by the transmitters. This allows the receiving lenslets 700 to also be densely packed, since the combination of lenslets 510, 511 and lenses 44,64 form a magnifying imaging system from the transducers in the tile 410 to the light entering the lenslet array 700. The two lens Fourier transform pairs create a compact optical system where light cones from one tile do not overlap the light cones from the adjacent tiles prior to exiting the lenslet array, when the tiles are densely packed, meaning that the distance between adjacent tiles is much smaller than the width of the tile—on the order of the distance between transmitters rather than the width of a tile.
As a specific example, lenslet 522 converts transmitted beamlet 150 into directed beamlet 152. Lenslet 524, along with lens 44 converts directed beamlet 152 into collimated beamlet 158. At the receiving end, lenslet 524 along with lens 64 converts collimated beamlet 158 into focused beamlet 164 which is received by receiver 621.
Thus transmitter 412 corresponds to receiver 621 (and vice versa in the case of a reverse operating system where transmitter 412 is a transmitter-receiver and receiver 621 is a receiver-transmitter).
With optoelectronic transducer 412 acting as a reflective modulator, the circularly polarized light 150 coming from 412 is directed 152 by lenslet 522 towards the center of lenslet 524. Lenslet 524 collimates directed beamlet 152 into collimated beamlet 158, at an angle substantially proportional to the position of the optoelectronic transducer 412 within tile 410. Lens 44 bends beamlet 158 toward the center of the optical train. Quarter-wave plate 40 converts the circularly polarized light in the beamlet to p-polarized light that is transmitted by polarizing beam splitters 31 and 30. Quarter-wave plate 60 converts the light in the beamlet to circular polarization and lens 64 bends this beamlet toward the center of the optical train. A lenslet pair in lenslet array 700 focuses the light onto optoelectronic transducer 621 in tile 620. Optoelectronic transducer 621 acting as an optical detector converts the signal from optical to electronic.
2C show the pattern of interconnectivity in two transverse dimensions. The horizontal dimension of
Returning to
Compact subarrays are enabled by the two lenslet Fourier transform. Two lenses separated by their focal length perform a Fourier transform of the light field. The lenslet nearest the modulator plane, 1530, steers the rays from each modulator/detector towards the center of a second lens, 1531. The second lens collimates these rays. This compact system allows one to densely pack subarrays at the input and output of a transposer, with minimal gaps between subarrays.
An optical transposer may be used in a neural network. We put a systolic-array matrix-multiplier on each subarray on one side of the transposer, and we put routing and memory on the other side of the transposer. This creates a connected set of systolic-array matrix-multipliers whose connections can be reconfigured on the routing and memory side. Memory registers are all a short distance from each systolic array due to the long range distributed connectivity of the transposer. This enables an efficient artificial neural network, since signal routing and memory access are the high latency, high energy dissipation parts of neural networks in current technology.
A system that moves blocks of data large distances by communicating blocks of data (tiles) with the use of a transpose, performing the same short-distance shift on each component of the dispersed block of data (itels), transposing back again so that the block of data (tile) is reconstructed with a large-distance shift proportional to the short-distance shift performed on each component.
Note that the numbering of elements is slightly different than in previous figures, because
Fourier transform elements 700 convert the position of the signals from fiber 62,300 to angles as shown in, for example,
Next the signals are transmitted to the terminating side of the device. Fourier transform elements 500 convert the angles of the arriving signals to positions to match them up with cores in terminating side multicore fiber 41,400. In
Returning to
The preceding describes just one optical pathway within one embodiment of the invention. Other pathways in the same embodiment connect additional pairs of originating transceivers to terminating transceivers. One pathway (in dotted lines) may couple light from originating transceiver 92350 to its terminating transceiver (not shown) on terminating tile 100050. Said light may couple through the same originating multicore fiber 62300, the same originating Fourier transform element 700, and the same array scale Fourier transform optics 65. For the transceiver 92350, the light will pass through a different terminating Fourier transform element 500B, and a different terminating multicore fiber 40050 such that it lands on a terminating transceiver placed on a different terminating tile 100050.
Another distinct pathway of the preceding embodiment may couple light emitted by a terminating transceiver 101123 back through the system such that the light illuminates an originating transceiver 92311.
Another distinct pathway of the preceding embodiment may start with light emitted from an originating transceiver on a different originating tile (not shown) such that the light passes through a different originating multicore fiber (not shown) in the originating array of multicore fibers (not shown). Said light will propagate through the array scale Fourier transform optics 65 such that it is transmitted through its terminating Fourier transform element 500 and then couples into a core within the terminating multicore fiber (40023) within the terminating array of multicore fibers, landing on a different terminating transceiver than the one labeled 101123 within the terminating tile 100023.
Another distinct pathway of the preceding embodiment may start with light emitted from an originating transceiver on a different originating tile (not shown) such that the light passes through a different originating multicore fiber (not labeled) in the originating array of multicore fibers (not labeled). Said light will propagate through the array scale Fourier transform optics 65 such that it is transmitted through a different terminating Fourier transform element (not labeled) within the terminating array of Fourier transform elements and then couples into a core within a different terminating multicore fiber (not labeled) within the terminating array of multicore fibers, landing on a different terminating transceiver within a different terminating tile.
In another embodiment, one or more originating tiles may be placed on one or more originating chip segments (not shown) which are distinct from the originating chip segment 90000 shown. In an extension of said embodiment, originating transceivers in originating tiles placed on the distinct originating chip segment may emit light which illuminates one terminating transceiver on the terminating tile 100023 shown here. In a separate extension of said embodiment, originating transceivers in originating tiles placed on the distinct originating chip segment may emit light which illuminates one terminating transceiver on a different terminating tile.
In another embodiment, one or more terminating tiles may be placed on one or more terminating chip segments which are distinct from the terminating chip segment 100000. In an extension of said embodiment, light emitted by one originating transceiver on the originating tile 92300 may illuminate terminating transceivers on the terminating tile placed on the distinct terminating chip segment. In another extension of said embodiment, light emitted by a different originating transceiver on a different originating tile 92300 may illuminate terminating transceivers on the terminating tile placed on the distinct terminating chip segment.
In one embodiment, light emitted by an originating transceiver may be emitted directly into one core of the multicore optical fiber in contact with the transceiver. In an alternative embodiment, the multicore optical fiber is not in contact, but is placed sufficiently close to couple light from the transceiver into one core of the multicore optical fiber. In an alternative embodiment, light emitted by an originating transceiver may be coupled into one core of the multicore optical fiber through a waveguide.
In one embodiment, light emitted by an originating transceiver may be coupled into one core of a multicore fiber using a lens. In one variation of said embodiment, the lens may be a bi-telecentric lens such as described by USPTO application 63/286,450 which is incorporated herein by reference. In one embodiment, light is coupled between one core of a terminating multicore fiber and the terminating transceiver using a bi-telecentric lens.
Throughout the preceding embodiments of the invention, a single optical pathway connects a single originating transceiver to a single terminating transceiver. Throughout the preceding embodiments, a single instance of array scale Fourier transform optics was used. Throughout the preceding embodiments, a single originating array of multicore fibers was used. Throughout the preceding embodiments, a single terminating array of multicore fibers was used. Throughout the preceding embodiments, a single originating array of Fourier transform elements was used. Throughout the preceding embodiments, a single terminating array of Fourier transform elements was used.
The variations of the preceding embodiments may be further combined with the following. In alternative embodiments of the invention, multicore fibers may be separated among many originating arrays of multicore fibers. In alternative embodiments of the invention, multicore fibers may be separated among many terminating arrays of multicore fibers. In alternative embodiments, multiple originating Fourier transform elements may be separated into multiple originating arrays of Fourier transform elements. In alternative embodiments, multiple terminating Fourier transform elements may be separated into multiple terminating arrays of Fourier transform elements.
The interconnection pattern is separable in each transverse dimension and the one-dimensional pattern is the same shuffle-exchange as shown
While the exemplary preferred embodiments of the present invention are described herein with particularity, those skilled in the art will appreciate various changes, additions, and applications other than those specifically mentioned, which are within the spirit of this invention. “Signals” can mean digital bits, digital on off, or phase and amplitude modulated signals, or analog. The optoelectronic transducer can be 2-mode semiconductor device, or polarization walk-off. The lenses 14,84,94,44,64, as shown all have same focal length, but can vary to expand beams or the like. Time multiplexing can be used for either transmit or receive or both. Optics for compact FT, tiles packing, transposer packing can be used.
The number of transmitters in an originating tile may be equal to the number of tiles in the destination or terminating array. The interconnection pattern between transmitters and receivers may be configured to be an N-way perfect shuffle. The transmitters may comprise LEDs or VCSELs.
| Number | Date | Country | |
|---|---|---|---|
| 62900225 | Sep 2019 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 17020661 | Sep 2020 | US |
| Child | 17929338 | US |
