CN116932459A - Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation - Google Patents
Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation Download PDFInfo
- Publication number
- CN116932459A CN116932459A CN202310965549.3A CN202310965549A CN116932459A CN 116932459 A CN116932459 A CN 116932459A CN 202310965549 A CN202310965549 A CN 202310965549A CN 116932459 A CN116932459 A CN 116932459A
- Authority
- CN
- China
- Prior art keywords
- optical
- signal
- optical waveguide
- input
- photon
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000004364 calculation method Methods 0.000 title claims abstract description 71
- 239000011159 matrix material Substances 0.000 title claims abstract description 23
- 230000003287 optical effect Effects 0.000 claims description 168
- 238000010521 absorption reaction Methods 0.000 claims description 25
- 238000006243 chemical reaction Methods 0.000 claims description 25
- 239000012782 phase change material Substances 0.000 claims description 23
- 238000000034 method Methods 0.000 claims description 18
- 230000008878 coupling Effects 0.000 claims description 16
- 238000010168 coupling process Methods 0.000 claims description 16
- 238000005859 coupling reaction Methods 0.000 claims description 16
- 230000008859 change Effects 0.000 claims description 11
- 230000005540 biological transmission Effects 0.000 claims description 9
- 238000013507 mapping Methods 0.000 claims description 6
- 239000002184 metal Substances 0.000 claims description 6
- 229910052751 metal Inorganic materials 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 4
- 230000031700 light absorption Effects 0.000 claims description 4
- 230000009471 action Effects 0.000 claims description 3
- 230000005284 excitation Effects 0.000 claims description 2
- 239000000969 carrier Substances 0.000 description 12
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 description 8
- 230000005693 optoelectronics Effects 0.000 description 8
- 229910052710 silicon Inorganic materials 0.000 description 8
- 239000010703 silicon Substances 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 4
- 238000005265 energy consumption Methods 0.000 description 4
- 239000000463 material Substances 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 239000004065 semiconductor Substances 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000004377 microelectronic Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000004806 packaging method and process Methods 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 1
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- 241001125929 Trisopterus luscus Species 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000002041 carbon nanotube Substances 0.000 description 1
- 229910021393 carbon nanotube Inorganic materials 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009792 diffusion process Methods 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005538 encapsulation Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000002347 injection Methods 0.000 description 1
- 239000007924 injection Substances 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000005468 ion implantation Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000012892 rational function Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000000758 substrate Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/76—Architectures of general purpose stored program computers
- G06F15/78—Architectures of general purpose stored program computers comprising a single central processing unit
- G06F15/7807—System on chip, i.e. computer system on a single chip; System in package, i.e. computer system on one or more chips in a single package
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
- G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
- G06F7/544—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices for evaluating functions by calculation
- G06F7/5443—Sum of products
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Hardware Design (AREA)
- Computing Systems (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
Abstract
The application discloses a photoelectric mixed calculation method and an array for on-chip large-scale matrix multiplication operation, which adopt an electric bus to sum current output signals with products of photoelectric mixed calculation units connected in parallel on different rows, so that the input power requirement of a light source is lower, and the wavelength requirement of input light of different rows is also reduced. Therefore, the photoelectric hybrid computing array reduces the requirement on a light source to a certain extent, and can realize the integrated application of the light source and the photon computing chip more quickly.
Description
Technical Field
The application relates to the technical field of photon calculation, in particular to a photoelectric hybrid calculation method and an array for on-chip large-scale matrix multiplication operation.
Background
As the semiconductor industry gradually enters the post-molar age, the development of integrated circuits continues to evolve in different directions. On the one hand, the development of new semiconductor materials, in particular carbon nanotubes, two-dimensional semiconductor materials, etc., continues the essence of Moore's law, continuing to shrink the device or chip size, i.e. "deep Moore"; on the other hand, for specific application fields, new architectures and heterogeneous integrated chips, such as neuromorphic chips, optoelectronic chips, quantum chips, etc., are developed, realizing "molar excess" (More than molar).
The photon chip based on the silicon-based photoelectronic technology integrates micron and nanometer scale photon, electron and photoelectronic devices on the same silicon substrate by using materials and technology compatible with an integrated circuit so as to realize the functional integration and advantage complementation of the microelectronic device and the photoelectronic device, obtain the photoelectronic chip with excellent performance, and is an effective way for solving the performance bottleneck and information congestion faced by the traditional integrated circuit. Thanks to the mature application of optical fiber communication, photons are used as information carriers, have more multiplexing dimensions, such as amplitude, phase, wavelength, mode, etc., than electrons, and further have larger bandwidth, faster speed and lower energy consumption. Early silicon-based photoelectronic chips are used to replace copper interconnection technology, solve the interconnection communication bottleneck between the processor core and the memory of microelectronic chip, and realize the microprocessor and the memory unit by microelectronic devices, while the photon devices mainly complete signal receiving and transmitting and information transmission. With the growing maturity of the silicon optical technology and the great advantage of optical communication, people pay attention to the silicon optical chip from information transmission to information processing, including the front application fields of analog calculation, quantum calculation, brain-like calculation and the like.
The realization technical path of the existing photon calculation chip is as follows: photon calculations are implemented by using a mach-zehnder interferometer (MZI) or a micro-ring structure (MMR), and the calculation modes are disclosed in, for example, chinese patents CN115905792 and CN113392965, and the calculation arrays formed based on these techniques are disclosed in, for example, chinese patents CN10407644 and CN116107037, but these techniques generally have the following disadvantages:
1. calculating array size limits: the photon calculation chip based on the MZI has larger area of a basic calculation unit, and when a silicon optical platform flows, a mask plate under the area of a conventional wafer is difficult to realize the application level of the calculation array scale.
2. The calculation speed is slow: the method mainly utilizes a thermo-optical tuning principle based on a silicon-based photoelectron technology, and simultaneously has the phenomenon of temperature drift (MMR structure is particularly remarkable) to influence the refractive index of the optical waveguide.
3. Design & encapsulation degree of difficulty is high: the photon calculation chip based on the MZI has more than 10 electrodes per basic calculation unit, the wiring quantity is extremely large, and the design difficulty and the packaging complexity are high.
The inventor firstly proposes a novel photon computing array which adopts a Crossbar matrix architecture (Crossbar); the cross switch matrix architecture is formed by crossing a group of input row waveguides which are parallel to each other and a group of output column waveguides which are parallel to each other, low-loss transmission of optical signals is realized by using the cross waveguides at the crossing points of the row waveguides and the column waveguides, and the outputs of photon calculation units on different rows are summed through column waveguide coupling devices on the column waveguides. Although the novel photon computing array performs summation through light, in engineering application, the cross waveguide and column waveguide coupling devices bring about larger loss, and the requirement on the whole input light source is higher. From the perspective of photon calculation field, the current integrated high-power light source has technical difficulties, and the light source and the chip cannot be integrated together in a short time, so that some photon calculation companies all need to use external high-power light sources.
Disclosure of Invention
The application aims to overcome the defects in the prior art and provides a photoelectric hybrid calculation method and an array for on-chip large-scale matrix multiplication operation.
The aim of the application is achieved by the following technical scheme:
an optoelectronic hybrid computing method for on-chip large-scale matrix multiplication operation, comprising the following steps:
s1, providing a photoelectric hybrid computing unit, wherein the photoelectric hybrid computing unit comprises a photon computing unit and a photoelectric conversion device which are connected with each other, and the photon computing unit generates an optical output signal with coded information; the photoelectric conversion device converts the light output signal into a current signal, and the magnitude of the current signal is proportional to the power of the light output signal;
s2, providing an electric bus, and summing corresponding current signals converted by a plurality of parallel photoelectric conversion devices to generate an electric output signal, wherein the sum of the current signals after addition is encoded in the electric output signal.
Preferably, the step S1 of "the photon calculating unit generates the optical output signal with encoded information" specifically includes:
s11, providing a write signal, and encoding a multiplier value into the write signal;
s12, mapping a multiplier value to a state of the photon calculation unit by using the writing signal, wherein the photon calculation unit comprises an optical waveguide and a modulation element optically coupled to the optical waveguide, and the modulation element modifies transmission, reflection, refraction or absorption characteristics of the optical waveguide according to the self state, wherein the self state of the modulation element can be adjusted by the writing signal; the state is represented by the absorption coefficient alpha or refractive index n of the optical waveguide to light;
s13, encoding the multiplicand value into an optical input signal of the photon calculation unit; the optical input signal is passed through the photon calculation unit to generate the optical output signal, wherein the product of the multiplier value and the multiplicand value is encoded in the optical output signal.
Preferably, the modulation element is an electro-optical modulator based on light absorption effect, the electro-optical modulator uses an electric signal as external excitation, and the absorption coefficient alpha of the optical waveguide containing free carriers to light is changed by injecting current or applying voltage in the doped region of the electro-optical modulator to change the concentration of the free carriers, so that the multiplication operation of the optical signal passing through the optical waveguide is realized.
Preferably, the modulation element is a phase change material deposited on the optical waveguide; the phase change material may optionally change its state by an optical signal as an external stimulus (i.e. a write signal), which is manifested by modifying the absorption coefficient α of light by an optical waveguide containing the phase change material.
Alternatively, the "the photon calculating unit generates the optical output signal with encoded information" in the step S1 is implemented using a mach-zehnder interferometer (MZI) or a micro-ring structure (MMR).
Preferably, the electrical bus is a metal interconnect layer on a photonic computing chip.
The application also discloses an optoelectronic hybrid computing array for on-chip large-scale matrix multiplication, comprising:
a set of photonic computing units comprising an optical waveguide and a modulating element optically coupled to the optical waveguide, the modulating element modifying a transmission, reflection, refraction or absorption characteristic of the optical waveguide according to the self state, wherein the self state of the modulating element is adjustable by a write signal; the state is represented by the absorption coefficient alpha or refractive index n of the optical waveguide to light; generating said optical output signal after the optical input signal passes through said photon calculation unit, wherein the product of a multiplier value and a multiplicand value is encoded in said optical output signal;
the photoelectric conversion devices are in one-to-one matching connection with the output sides of the optical waveguides of the photon calculation unit and are used for converting the optical output signals into current signals, and the magnitude of the current signals is in direct proportion to the power of the optical output signals;
a crossbar matrix architecture; the cross switch matrix architecture is formed by crossing a group of mutually parallel input optical waveguides and a group of mutually parallel electric buses; the optical input signal is transmitted through the input optical waveguide;
the photonic computing units are present at the intersection of each input optical waveguide and the electrical bus in the crossbar architecture, the input side of the optical waveguide of each photonic computing unit being coupled to an adjacent input optical waveguide by a waveguide coupling device that evanescently couples a portion of the optical power from the adjacent input optical waveguide, wherein the coupled optical power depends on the length of the portion of the waveguide coupling device that is positioned adjacent to and extends parallel to the adjacent input optical waveguide;
and each electric bus is connected with a plurality of parallel photoelectric conversion devices, sums corresponding current signals converted by the plurality of parallel photoelectric conversion devices to generate electric output signals, and the sum of the current signals after addition operation is encoded in the electric output signals.
Preferably, the input optical waveguide and the electrical bus are perpendicular to each other.
Preferably, the wavelengths of the light input signals of the input light waveguides of the different rows are different or the same.
Preferably, the optical input signal is used for evenly distributing the optical power of the optical input signal to each of the photon computing units on the same row under the action of the waveguide coupling device.
Preferably, the number of columns of the input optical waveguide and the electric bus is greater than or equal to 2.
Preferably, the electrical bus is a metal interconnect layer on the photonic computing chip.
The beneficial effects of the application are mainly as follows:
(1) The area of the basic calculation unit is smaller, and the mask plate under the conventional wafer area can realize the application level of the calculation array scale; the modulation speed is high, the modulation speed can be increased to nanosecond level through electric pulse, so that the calculation performance of the photon calculation chip is improved, and meanwhile, the phenomenon of temperature drift is avoided; each basic calculation unit only has 2 electrodes, the wiring quantity is small, and the design difficulty and the packaging complexity are low;
(2) The photoelectric hybrid computing array provided by the application adopts the photoelectric conversion device and the electric bus to sum the optical output signals with products of the photon computing units on different rows, firstly, the input power requirement on the light source is lower, and secondly, the wavelength requirement on the input light of different rows is also reduced, so that the photoelectric hybrid computing array reduces the requirement on the light source to a certain extent, and the integrated application of the light source and the photon computing chip can be realized more quickly.
Drawings
The technical scheme of the application is further described below with reference to the accompanying drawings:
fig. 1: the flow diagram of the photoelectric hybrid calculation method is shown in the specification;
fig. 2: schematic diagram of a preferred embodiment of the optoelectronic hybrid computing array of the present application;
fig. 3: a schematic diagram of a preferred embodiment 3*3 of the present application is a photoelectric hybrid computing array;
fig. 4: a schematic diagram of a second embodiment of the opto-electronic hybrid computing array of the present application;
fig. 5: schematic diagram of a third embodiment of the opto-electronic hybrid computing array of the present application;
fig. 6: a schematic structural diagram of a photon calculation unit according to a preferred embodiment of the present application;
fig. 7: the working principle of the photon calculation unit in fig. 6 for multiplication calculation is schematically shown.
Detailed Description
The present application will be described in detail below with reference to specific embodiments shown in the drawings. The embodiments are not limited to the present application, and structural, methodological, or functional modifications of the application from those skilled in the art are included within the scope of the application.
The application will be described in detail below with reference to the drawings in connection with embodiments.
The application provides an optoelectronic hybrid computing array for large-scale matrix multiplication on a chip, which has a specific structure shown in a preferred embodiment in fig. 2 and comprises an optoelectronic hybrid computing unit, a waveguide coupling device 5, an input optical waveguide 31 and an electric bus 32. The photoelectric hybrid computing unit includes a photon computing unit 1 and a photoelectric conversion device 2.
The photon calculation unit 1 is adopted to carry out multiplication operation, and comprises an optical waveguide 11 and a modulation element 12 optically coupled to the optical waveguide, wherein the modulation element modifies the transmission, reflection, refraction or absorption characteristics of the optical waveguide according to the self state, and the self state of the modulation element can be adjusted through writing signals; the state is represented by the absorption coefficient α or refractive index n of the optical waveguide for light. The optical input signal is passed through the photon calculation unit to generate the optical output signal, wherein the product of the multiplier value and the multiplicand value is encoded in the optical output signal.
In the present application, the modulation element 12 includes three forms.
In the preferred embodiment shown in fig. 2, the modulating element 12 (a) is an electro-optic modulator based on the light absorption effect, which uses an electrical signal as an external stimulus to change the absorption coefficient α of the optical waveguide containing free carriers by injecting a current or applying a voltage to change its free carrier concentration in its doped region, so that the optical signal passing through the optical waveguide is absorbed by the free carriers to implement a multiplication operation.
The specific multiplication process is shown in fig. 6 and 7, and comprises the following steps:
s11, providing a write signal, and encoding a multiplier value into the write signal 108;
s12 mapping a multiplier value to a state of a doped region of the modulation element 12 using the write signal 108, the state representing an absorption coefficient α of the optical waveguide for light at different free carrier concentrations;
s13, encoding the multiplicand value to an optical input signal 106 of the photon calculation unit;
s14, the optical input signal 106 passes through the optical absorption region of the modulation element 12 via the optical waveguide 11, generating an optical output signal 107, wherein the product of the multiplier value and the multiplicand value is encoded in the optical power of the optical output signal.
Wherein a write signal 108 for modulating the free carrier concentration in the element 12 is provided by the electrical signal generator 105. The two electrical interconnection devices in the metal layer of the electrical signal generator 105 transmit the generated electrical signal to the doped regions 103a, 103b of the modulation element 12 via the first contact electrode 104a, the second contact electrode 104b, respectively. The electric signal causes the free carriers to directionally move under the action of the electric field, and the change of parameters of the electric signal can change the concentration of the free carriers in the doped region, so that the absorption coefficient of the optical waveguide containing the free carriers to light is changed. The electro-optical modulator based on the light absorption effect comprises a doped region based on semiconductor doping technology such as ion implantation or high-temperature diffusion and a pair of contact electrodes which form ohmic contact or Schottky contact with the doped region.
The electrical signal generator 105 uses an external signal, i.e. the write signal 108 to map the multiplier value b to the absorption coefficient α of the optical waveguide containing free carriers to light, and the modulating element 12 injects a current or applies a voltage to change the concentration of free carriers (charge or holes) in the doped region, changing the absorption coefficient α of the optical waveguide containing free carriers to light. The energy Pwrite of the external electrical signal has a mapping in the form of a elementary or rational function with the absorption coefficient α of the light, the optical input signal 106 decays inversely proportional to the absorption coefficient α of the light with the optical waveguide containing free carriers, resulting in a multiplication output signal 107 corresponding to the external electrical signal write signal 108 and the optical input signal 106, the output signal pout=α×pin of the optical waveguide 11 being the result of mapping the multiplier value b to α and the multiplicand value a to Pin.
Emerging photonic devices based on phase change materials have been widely studied in recent years. The phase-change material has the advantages of high reading and writing speed (nanosecond magnitude) and high circulation times>10 12 ) The method has the characteristics of low power consumption and the like, can be compatible with the existing CMOS process, and has low technical realization difficulty and industrial cost. The phase change material has larger difference of optical and electrical characteristics in crystalline state and amorphous state, can induce phase change in various modes such as heat, light, electricity and the like, and has stable characteristics. The phase change material does not need to maintain static bias voltage for light modulation, and can maintain a certain state unchanged at room temperature. Therefore, the theoretical static control energy consumption of the phase change material is zero, the energy consumption of the system is greatly reduced, the stability of the system is also improved, and the characteristics enable the phase change material to be potential as a basic functional material for photon calculation.
In a second embodiment, as shown in fig. 4, the modulating element 12 (b) employs a phase change material 12' deposited on an optical waveguide; the phase change material 12 'uses an optical signal as an external stimulus (i.e., a write signal) to change its state, which is manifested by modifying the absorption coefficient α of light by an optical waveguide containing the phase change material 12'. The phase-change material is selected as the modulation element, and has the advantages that the phase-change material is nonvolatile, can be kept in the state after being subjected to modulation power failure, and is extremely suitable for the application scene of artificial intelligence for reasoning. Because the phase change material has the characteristic of non-volatility, compared with the technical path of the existing photon calculation chip, the phase change material does not maintain the generation of power consumption and has extremely high calculation power energy consumption ratio.
Specifically, an optical signal, for example, a specific high-power optical signal is transmitted through the optical waveguide, and the state of the phase change material itself can be changed by adding a specific time, that is, the absorption coefficient α of the optical waveguide containing the phase change material to light is modified. The photoelectric hybrid computing array based on the phase change material reduces the requirement on a light source to a certain extent, and can realize the integrated application of the light source and the photon computing chip more quickly.
In a third embodiment, as shown in fig. 5, the modulating element (c) is a modulator based on the phase change material. When modulated with an optical signal, there is no doped region. When modulated with an electrical signal, a doped region is included. An optical signal or an electrical signal may be selectively employed as the write signal after employing this embodiment. The multiplication process using the electrical signal for modulation is similar to the preferred embodiment, and the multiplication process using the optical signal for modulation is similar to the second embodiment, and therefore will not be described again.
Of course, the multiplication of the present application can also be implemented using Mach-Zehnder interferometers (MZIs) or micro-Ring Structures (MMRs) of the prior art.
The photoelectric conversion device 2 is connected with the output side of the optical waveguide of the photon calculation unit 1 in a matching way, and converts the optical output signal into a current signal, wherein the magnitude of the current signal is proportional to the power of the optical output signal. Therefore, the number and position settings of the photoelectric conversion devices 2 are in one-to-one correspondence with the photon calculation units 1.
The photoelectric hybrid computing array adopts a cross switch matrix architecture (Crossbar); formed by the intersection of a set of input optical waveguides 31, which may be doped, and a set of said electrical buses 32. The optical input signal is transmitted through the input optical waveguide 31. The input optical waveguides 31, which may be doped, are arranged in parallel between rows and the electrical buses 32 are arranged in parallel between columns. Preferably, the input optical waveguide which can be doped is perpendicular to the electrical bus. In the present application, the number of columns of the input optical waveguide 31 and the electrical bus 32, which can be doped, is 2 or more.
There is one of the photonic computing units 1 at the intersection of each input optical waveguide 31 and the electrical bus 32 in the crossbar architecture, the input side of the optical waveguide 11 of each of the photonic computing units 1 being coupled to an adjacent input optical waveguide 31 by a waveguide coupling device 5, the waveguide coupling device 5 evanescently coupling a portion of the optical power from the adjacent input optical waveguide 31, wherein the coupled optical power depends on the length of the portion of the waveguide coupling device 5 that is positioned adjacent to and extends parallel to the adjacent input optical waveguide.
The other side of the optical waveguide 11 is coupled to an adjacent electrical bus 32 via the photoelectric conversion device 2. Preferably, the electrical bus 32 employs a metal interconnect layer on a photonic computing chip.
Each of the electrical buses 32 is connected with a plurality of parallel photoelectric conversion devices 2, and the electrical buses 32 sum corresponding current signals converted by the plurality of parallel photoelectric conversion devices 2 to generate electrical output signals, and the sum of the current signals after addition is encoded in the electrical output signals.
The application is characterized in that a Crossbar (Crossbar) architecture is adopted to realize large-scale matrix multiplication and addition operation. Namely: after multiplication by a photon calculation unit, the optical output signals with products are converted into current signals through a photoelectric conversion device, and then summation is carried out on an electric bus.
In a preferred embodiment, the wavelengths of the light input signals of the input row waveguides 31 of different rows are different, so as to avoid interference phenomenon generated by the light during addition operation, and influence the accuracy of photon calculation. Because the application adopts the photoelectric hybrid computing architecture based on the electric bus, the wavelengths of the light of the input row waveguides 31 of different rows can be the same, and the requirement on the wavelength of a light source is reduced.
As shown in fig. 1 and 2, the present application provides a method for performing matrix multiplication and addition in an optical domain, which simply includes the following steps:
s1, providing a photoelectric hybrid computing unit, wherein the photoelectric hybrid computing unit comprises a photon computing unit and a photoelectric conversion device which are connected with each other, and the photon computing unit generates an optical output signal with coded information; the photoelectric conversion device converts the light output signal into a current signal, and the magnitude of the current signal is proportional to the power of the light output signal;
s2, providing an electric bus, and summing corresponding current signals converted by a plurality of parallel photoelectric conversion devices to generate an electric output signal, wherein the sum of the current signals after addition is encoded in the electric output signal.
The method is used for calculating an m×n-order matrix p×u=a:
the method specifically comprises the following steps:
(1) Encoding the weight matrix U into the write signal, for example, inputting the code U11 into the first row and first column of photon calculation units, inputting the code U12 into the first row and second column of photon calculation units, inputting the code U21 into the second row and first column of photon calculation units, and so on;
(2) Mapping the multiplier value to a state of the modulation element 12 or 12' in each photon calculation unit using a write signal, the state representing an absorption coefficient α of the light guide 11 for light;
(3) Encoding the input data matrix P into an optical input signal whose optical power is equally distributed to each of the cells on the same row of input optical waveguides 31 by the waveguide coupling device 5; for example, the code P1 is input into the input optical waveguide of the first row, the code P2 is input into the input optical waveguide of the second row, and so on;
(4) The optical input signal passes through the optical waveguide 11 and then passes through the modulating element of the photon calculating unit to generate an optical output signal, wherein the product of the multiplier value and the multiplicand value is encoded in the optical output signal, the optical power of the optical output signal converts the optical signal into a current signal through the photoelectric conversion device and is converged on the electric bus for summation, the magnitude of the current signal is in direct proportion to the power of the optical output signal, the optical output signal converts the current signal into a current signal response beta, and in practical application, the value of beta depends on the technological level of a silicon optical chip manufacturing plant.
For example: the power of the optical output signal of the first column is:
r 11 d 11 P 1 U 11 +r 21 d 21 P 2 U 21 +……+r m1 d m1 P m U m1 ＝A 1
the common factor in the above formula is beta/n. And so on.
Fig. 3 discloses a schematic diagram of an embodiment 3*3 of a hybrid photovoltaic computing array, implementing a computing third order matrix P x u=a,
wherein:
101-1, 101-2, 101-3: an input optical waveguide;
102-1, 102-2, 102-3: an electrical bus;
p1, P2, P3: inputting (reading) a signal;
a1, A2, A3: an electrical output signal;
u11, U12, U13, U21, U22, U23, U31, U32, U33: a photon calculation unit for multiplication;
r11, r12, r13, r21, r22, r23, r31, r32, r33: a waveguide coupling device;
d11, d12, d13, d21, d22, d23, d31, d32, d33: a photoelectric conversion device;
the method according to the application can be implemented to calculate the optical power output of the first column as:
the common factor in the above formula is beta/3.
Compared with a photon calculation array based on optical bus summation, the photoelectric hybrid calculation array based on the electric bus summation has no loss caused by a cross waveguide and a column waveguide coupling device, and meanwhile, the requirement on the wavelength of input light is lower because of no cross transmission structure in an optical path, and all input channels can be of the same wavelength or different wavelengths.
The injection current or the applied voltage is mainly in the form of electric pulse, the duration of the electric pulse is at nanosecond level, and meanwhile, the duration of the electric pulse determines the modulation speed of the photoelectric hybrid computing array, so that the computing performance of the photoelectric hybrid computing array is determined; and meanwhile, the phenomenon of temperature drift is avoided.
The above list of detailed descriptions is only specific to practical embodiments of the present application, and they are not intended to limit the scope of the present application, and all equivalent embodiments or modifications that do not depart from the spirit of the present application should be included in the scope of the present application.
Claims (12)
1. The photoelectric hybrid calculation method for the on-chip large-scale matrix multiplication operation is characterized by comprising the following steps of:
the method comprises the following steps:
s1, providing a photoelectric hybrid computing unit, wherein the photoelectric hybrid computing unit comprises a photon computing unit and a photoelectric conversion device which are connected with each other, and the photon computing unit generates an optical output signal with coded information; the photoelectric conversion device converts the light output signal into a current signal, and the magnitude of the current signal is proportional to the power of the light output signal;
s2, providing an electric bus, and summing corresponding current signals converted by a plurality of parallel photoelectric conversion devices to generate an electric output signal, wherein the sum of the current signals after addition is encoded in the electric output signal.
2. The method of optoelectric hybrid calculation of claim 1, wherein: the step S1 of "the photon calculating unit generates the optical output signal with encoded information" specifically includes:
s11, providing a write signal, and encoding a multiplier value into the write signal;
s12, mapping a multiplier value to a state of the photon calculation unit by using the writing signal, wherein the photon calculation unit comprises an optical waveguide and a modulation element optically coupled to the optical waveguide, and the modulation element modifies transmission, reflection, refraction or absorption characteristics of the optical waveguide according to the self state, wherein the self state of the modulation element can be adjusted by the writing signal; the state is represented by the absorption coefficient alpha or refractive index n of the optical waveguide to light;
s13, encoding the multiplicand value into an optical input signal of the photon calculation unit; the optical input signal is passed through the photon calculation unit to generate the optical output signal, wherein the product of the multiplier value and the multiplicand value is encoded in the optical output signal.
3. The method of optoelectric hybrid calculation of claim 2, wherein: the modulation element is an electro-optical modulator based on light absorption effect, the electro-optical modulator uses an electric signal as external excitation, and the free carrier concentration of the electro-optical modulator is changed by injecting current or applying voltage in a doped region of the electro-optical modulator so as to change the absorption coefficient alpha of the optical waveguide containing the free carrier to light, thereby realizing multiplication operation of the optical signal passing through the optical waveguide.
4. The method of optoelectric hybrid calculation of claim 2, wherein: the modulation element is a phase change material deposited on the optical waveguide; the phase change material may optionally change its state by an optical signal as an external stimulus (i.e. a write signal), which is manifested by modifying the absorption coefficient α of light by an optical waveguide containing the phase change material.
5. The method of optoelectric hybrid calculation of claim 1, wherein: the "the photon calculation unit generates the optical output signal with encoded information" in the step S1 is implemented using a mach-zehnder interferometer (MZI) or a micro-ring structure (MMR).
6. The method of optoelectric hybrid calculation of claim 1, wherein: the electrical bus is a metal interconnect layer on the photonic computing chip.
7. The photoelectric hybrid computing array for on-chip large-scale matrix multiplication operation is characterized in that: the system comprises a group of photon calculation units, a light source module and a light source module, wherein the photon calculation units comprise an optical waveguide and a modulation element optically coupled to the optical waveguide, the modulation element modifies the transmission, reflection, refraction or absorption characteristics of the optical waveguide according to the self state, and the self state of the modulation element can be adjusted through a write signal; the state is represented by the absorption coefficient alpha or refractive index n of the optical waveguide to light; generating said optical output signal after the optical input signal passes through said photon calculation unit, wherein the product of a multiplier value and a multiplicand value is encoded in said optical output signal;
the photoelectric conversion devices are in one-to-one matching connection with the output sides of the optical waveguides of the photon calculation unit and are used for converting the optical output signals into current signals, and the magnitude of the current signals is in direct proportion to the power of the optical output signals;
a crossbar matrix architecture; the cross switch matrix architecture is formed by crossing a group of mutually parallel input optical waveguides and a group of mutually parallel electric buses; the optical input signal is transmitted through the input optical waveguide;
the photonic computing units are present at the intersection of each input optical waveguide and the electrical bus in the crossbar architecture, the input side of the optical waveguide of each photonic computing unit being coupled to an adjacent input optical waveguide by a waveguide coupling device that evanescently couples a portion of the optical power from the adjacent input optical waveguide, wherein the coupled optical power depends on the length of the portion of the waveguide coupling device that is positioned adjacent to and extends parallel to the adjacent input optical waveguide;
and each electric bus is connected with a plurality of parallel photoelectric conversion devices, sums corresponding current signals converted by the plurality of parallel photoelectric conversion devices to generate electric output signals, and the sum of the current signals after addition operation is encoded in the electric output signals.
8. The electro-optical hybrid computing array of claim 7, wherein: the input optical waveguide and the electrical bus are perpendicular to each other.
9. The electro-optical hybrid computing array of claim 7, wherein: the wavelengths of the light input signals of the input light waveguides of the different rows are different or the same.
10. The electro-optical hybrid computing array of claim 7, wherein: the optical input signals are used for evenly distributing the optical power of the optical input signals to each photon computing unit on the same row under the action of the waveguide coupling device.
11. The electro-optical hybrid computing array of claim 7, wherein: the number of columns of the input optical waveguide and the electric bus is more than or equal to 2.
12. The electro-optical hybrid computing array of claim 7, wherein: the electrical bus is a metal interconnect layer on the photonic computing chip.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310965549.3A CN116932459A (en) | 2023-08-02 | 2023-08-02 | Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310965549.3A CN116932459A (en) | 2023-08-02 | 2023-08-02 | Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116932459A true CN116932459A (en) | 2023-10-24 |
Family
ID=88377101
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310965549.3A Pending CN116932459A (en) | 2023-08-02 | 2023-08-02 | Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116932459A (en) |
-
2023
- 2023-08-02 CN CN202310965549.3A patent/CN116932459A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Lentine et al. | Evolution of the SEED technology: bistable logic gates to optoelectronic smart pixels | |
| CN114158277B (en) | Photoelectric Computing Platform | |
| CN113805641B (en) | Photonic neural network | |
| CN111538368B (en) | Photon information processing chip | |
| CN116736933A (en) | Photon computing array for on-chip large-scale matrix multiplication | |
| CN113644984A (en) | Optical logic element for photoelectric digital logic operation and logic operation method thereof | |
| CN116932459A (en) | Photoelectric hybrid calculation method and array for on-chip large-scale matrix multiplication operation | |
| Kirk et al. | Parallel optoelectronic data transcription with fan-out between planes of PnpN optical thyristors | |
| Nitta et al. | Optical neurochip with learning capability | |
| CN114707629A (en) | Matrix calculation method based on light scattering and optical neural network | |
| Cong et al. | Experimental Demonstration of XOR Separation by On-chip Training a Linear Silicon Photonic Circuit | |
| Zhang et al. | TeMPO: Efficient Time-Multiplexed Dynamic Photonic Tensor Core for Edge AI with Compact Slow-Light Electro-Optic Modulator | |
| CN113839715A (en) | Optical neural network and chip | |
| WANG et al. | Development and thinking of silicon photonics heterogenous integration | |
| CN117130099A (en) | Ultra-large scale photoelectric hybrid calculation array based on multimode interferometer | |
| Li et al. | Lateral separate absorption charge multiplication Ge-on-Si avalanche photodiode with low dark current in linear mode | |
| CN117234276A (en) | Hybrid photon computing unit, computing method and computing array for on-chip large-scale matrix multiplication operation | |
| Tossoun et al. | Heterogeneously Integrated III–V on Silicon Photonics for Neuromorphic Computing | |
| CN116449631B (en) | Multi-waveguide integrated logic gate device, photoelectric device and preparation method thereof | |
| CN116107103B (en) | High resolution vertical optical phase modulation array | |
| Peserico et al. | Emerging Devices and Packaging Strategies for Electronic-Photonic AI Accelerators | |
| CN117896011A (en) | Method for performing operation in optical domain by photon calculation unit | |
| Nitta et al. | Monolithic integration of optical neurochip with variable sensitivity photodetector | |
| WO2023140788A2 (en) | Photonic neural network accelerator | |
| Barnes et al. | Parallel interconnection and neural network systems employing semiconductor array technology and computer-generated holography |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |