CN109981172B - All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay - Google Patents
All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay Download PDFInfo
- Publication number
- CN109981172B CN109981172B CN201910154200.5A CN201910154200A CN109981172B CN 109981172 B CN109981172 B CN 109981172B CN 201910154200 A CN201910154200 A CN 201910154200A CN 109981172 B CN109981172 B CN 109981172B
- Authority
- CN
- China
- Prior art keywords
- wavelength
- dispersion
- chip
- modulation
- optical fiber
- 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.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B10/00—Transmission systems employing electromagnetic waves other than radio-waves, e.g. infrared, visible or ultraviolet light, or employing corpuscular radiation, e.g. quantum communication
- H04B10/25—Arrangements specific to fibre transmission
- H04B10/2507—Arrangements specific to fibre transmission for the reduction or elimination of distortion or dispersion
- H04B10/2513—Arrangements specific to fibre transmission for the reduction or elimination of distortion or dispersion due to chromatic dispersion
- H04B10/25133—Arrangements specific to fibre transmission for the reduction or elimination of distortion or dispersion due to chromatic dispersion including a lumped electrical or optical dispersion compensator
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04J—MULTIPLEX COMMUNICATION
- H04J14/00—Optical multiplex systems
- H04J14/02—Wavelength-division multiplex systems
- H04J14/0227—Operation, administration, maintenance or provisioning [OAMP] of WDM networks, e.g. media access, routing or wavelength allocation
- H04J14/0241—Wavelength allocation for communications one-to-one, e.g. unicasting wavelengths
- H04J14/0242—Wavelength allocation for communications one-to-one, e.g. unicasting wavelengths in WDM-PON
Abstract
An all-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay adopts wavelength division multiplexing signals as light sources, and weights modulation signals in a wavelength domain through a cascaded wavelength selective modulator after broadband modulation or weights modulation signals in a time domain through a cascaded broadband modulator so as to realize multiplication; then, the weighted optical signals pass through a dispersion medium, so that dispersion delay superposition is carried out on each wavelength, and the addition operation of the signals is realized; the invention adopts the multiplexing technology to realize two linear weighting modes, further realizes the multidimensional flexible regulation and control of the multiply-add unit, and greatly improves the expandability of the system.
Description
Technical Field
The invention relates to a technology in the field of optical computing, in particular to a chip system for realizing high-speed large-scale matrix multiply-add operation and further realizing an integrated photonic neural network.
Background
Linear multiply-add operation, as a basic unit of an Artificial Neural network (ans), is one of the research hotspots in recent years in the field of optical computing. However, the existing photoelectric multiply-add operation scheme neglects the positioning of the computing hardware in the whole computer system, and is mainly reflected in: firstly, the expandability of the optical computing structure provided at present is poor, and some designs can only solve the matrix vector multiplication of a fixed order and cannot meet the requirement of large-scale data operation; secondly, the potential for its use is very limited due to the difficulty of small scale integration of some optical components (e.g., erbium doped fiber amplifiers). The parallel computing hardware field represented by neural network acceleration hardware urgently needs a novel optical computing technology to break through from the aspects of mechanism and method, meet the requirements of high speed, high bandwidth and low power consumption, and realize the linear all-optical matrix multiply-add operation with strong expandability.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides an all-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay, two linear weighting modes are implemented by adopting a multiplexing technology, so that multidimensional flexible regulation and control of a multiply-add unit are realized, and the expandability of a system is greatly improved.
The invention is realized by the following technical scheme:
the invention relates to a method for realizing multiplication and addition of an all-optical matrix based on multi-wavelength modulation and dispersion time delay, which adopts wavelength division multiplexing signals as light sources, weights modulation signals in a wavelength domain through a cascade wavelength selective modulator after broadband modulation or weights modulation signals in a time domain through a cascade broadband modulator so as to realize multiplication operation; and then, the weighted optical signals pass through a dispersion medium, so that dispersion delay superposition is carried out on each wavelength, and the addition operation of the signals is realized.
The wavelength selective modulator is characterized in that: a device that individually modulates a signal for a particular wavelength. The optical fiber has a small free spectrum range, so that the optical fiber has strong wavelength dependence and wavelength selectivity. Such as a micro-ring modulator, is common.
The broadband modulator is characterized in that: devices that commonly modulate multiple wavelength signals may be considered wavelength insensitive due to their large free spectral range. Common multi-wavelength modulators are Mach-Zehnder modulators.
The dispersion delay refers to: the transmission rates of the multi-wavelength optical signals in the waveguide are different, so that the time for the multi-wavelength signals simultaneously entering the waveguide to reach a receiving end is different, and time delay exists among different wavelengths.
The invention relates to an on-chip integrated system for realizing the method, which comprises a light source, an on-chip weighting device, a dispersion medium, a photoelectric detector and a waveform generator which are connected in sequence, wherein: each wavelength of the wavelength division multiplexing light source carries a signal to be weighted after broadband modulation, multiplication is realized through wavelength domain or time domain weighting, then corresponding time delay is carried out on each weighted wavelength signal by a designed dispersion medium, and addition operation is realized on a photoelectric detector.
The light source is a 1550nm waveband multi-wavelength light source and is realized by generating a light frequency comb on a chip through a micro-ring resonant cavity and a nonlinear effect.
The on-chip weighting device is a micro-ring modulator serving as an on-chip wavelength domain weighting device or a Mach-Zehnder modulator serving as a time domain weighting device.
And a polarization controller and a wave former are preferably further arranged between the on-chip weighter and the light source.
The waveform generator is a pulse waveform generator or an arbitrary waveform generator with an electric amplifier.
The dispersion medium is a dispersion compensation optical fiber, and the dispersion coefficient of the dispersion medium is-160 to-136 ps/nm/km.
The invention relates to the application of the integrated system on a chip, which is used as an operation unit of a photon neural network and forms a feedforward neural network structure through a plurality of cascades.
The application further realizes the nonlinear activation function of the neural network through the multiplication operation or the addition operation.
The artificial neural network algorithm involves a large number of matrix-vector multiplication operations, and by taking the method provided by the invention as the most basic operation unit and according to the logic structure of the artificial neural network algorithm, a linear part of the artificial neural network algorithm except a non-linear activation function can be built.
Technical effects
Compared with the prior art, the invention realizes multidimensional flexible weighting and operation through a multiplexing technology, realizes high-speed and large-scale matrix operation on a single weighting system, and greatly reduces the complexity of the system. The invention realizes the autocorrelation operation of the 7-bit m sequence in a WDW weighting mode, and the operation rate can reach 1.18 multiplied by 1011Sub-multiply-add operations per second (MAC/s); experimentally at 2.69X 10 by TDW weighting9The rate of MAC/s realizes 4-order matrix vector multiplication operation, and is 5 multiplied by 10 in simulation8The operation rate of MAC/s enables edge extraction of binary handwriting images of size 32 × 32.
Drawings
FIG. 1 is a schematic diagram of the principles of the present invention;
in the figure: WB: weight coefficient matrix, DM: a dispersive medium;
FIG. 2 is a weight coefficient matrix for Wavelength Domain Weighting (WDW) and Time Domain Weighting (TDW) according to the present invention;
in the figure: (a) a weight coefficient matrix of a wavelength domain weighting method, and (b) a weight coefficient matrix of a time domain weighting method;
FIG. 3 is a block diagram of an implementation of a system;
FIG. 4 is a schematic diagram of the present invention for implementing the autocorrelation operation of a 7-bit m-sequence by wavelength domain weighting;
FIG. 5 is a block diagram of an experiment and results of autocorrelation calculations;
in the figure: (a) experimental setup, (b) non-equidistant optical frequency comb as multi-wavelength light source, (c) experimental and simulation results of 7-bit m-sequence autocorrelation operation;
FIG. 6 is a schematic diagram of the 4 th-order matrix vector multiplication implemented by Time Division Multiplexing (TDM) -assisted time domain weighting according to the present invention;
FIG. 7 is the experimental block diagram and the result of the 4 th-order matrix vector multiplication
In the figure: (a) setting an experiment, (b) forming a 4-order matrix and a vector time domain waveform, (c) calculating a result waveform;
FIG. 8 is a simulation block diagram and results of edge extraction for a 32 × 32 size handwritten binary image by time domain weighting;
in the figure: (a) setting a simulation system, (b) setting a binary handwriting image to be extracted, (c) extracting a binary handwriting image after edges, and (d) changing the influence of the length of a dispersion compensation optical fiber on the extraction effect.
Detailed Description
Example 1
As shown in fig. 5, the present embodiment relates to a system on chip for implementing a wavelength domain weighting to implement a 7-bit m-sequence autocorrelation operation, including: the device comprises a multi-wavelength optical frequency comb, a waveform shaper, a Mach-Zehnder modulator, a dispersion compensation optical fiber, a photoelectric detector and a pulse waveform generator which are connected in sequence.
In the embodiment, the multi-wavelength light source is an unequal interval optical frequency comb, the intervals are respectively 36GHz and 108GHz, the dispersion coefficient of the dispersion compensation optical fiber is-160 ps/nm/km, the length is 1.29km, the pulse waveform generator generates a 7-bit m sequence of 16.875Gb/s, and a 6-bit isolation code element is additionally arranged to prevent the dispersion from delaying and then calculate the time slot overlapping coverage. Modulation speedThe ratio f, the wavelength interval delta lambda, the dispersion coefficient D and the optical fiber length l satisfy: 1/f is D × Δ λ × l. The experimental waveform obtained by the photoelectric detector is consistent with the simulated waveform, an obvious correlation peak can be seen, and the interval between the two correlation peaks is 770 ps. The calculated speed is: 16.875/13X 7X 13X 109≈1.18×1011MAC/s。
Example 2
As shown in fig. 7, the present embodiment relates to a system on chip for implementing time domain weighting assisted by time division multiplexing to implement 4-order matrix vector multiply-add, including: the optical fiber dispersion compensation device comprises a four-channel multi-wavelength light source, two cascaded Mach-Zehnder modulators, a dispersion compensation optical fiber, a photoelectric detector, an arbitrary waveform generator, an electric amplifier and an oscilloscope which are sequentially connected.
In the embodiment, the wavelengths of the multi-wavelength light source are 1550.12nm, 1553.12nm, 1556.12nm and 1559.12nm respectively; the arbitrary waveform generator generates a 4 × 4 matrix H and a 4 × 1 vector x of 5.37Gb/s in a time division multiplexed manner. The dispersion compensation fiber had a dispersion coefficient of-136 ps/nm/km and a length of 455 m. The result of the operation occurs at an intermediate time in the time division multiplex slot. The calculated speed is: 5.37/8X 4X 109≈2.69×1011MAC/s。
Example 3
As shown in fig. 8, the present embodiment relates to a simulation system for realizing handwriting edge extraction of a 32 × 32 binary handwriting image, comprising: the optical fiber dispersion compensation device comprises multiple wavelengths, two cascaded Mach-Zehnder modulators, a dispersion compensation optical fiber, a photoelectric detector, an arbitrary waveform generator, an electric amplifier and an oscilloscope which are sequentially connected.
In the embodiment, the light source is 9 wavelengths with 1550nm band wavelength interval of 100GHz, the two-dimensional image and the 3 × 3 laplacian are converted into a one-dimensional data stream in parallel-serial conversion, wherein the laplacian is represented in a 5-bit binary form, and two cascaded mach-zehnder modulators are generated by an arbitrary waveform generator in a time division multiplexing manner and respectively modulate at the rates of 1Gb/s and 5 Gb/s. The dispersion coefficient of the dispersion compensating fiber was set to-160 ps/nm/km, and the length could be calculated to be 7.8125 km. The calculation result received by the photoelectric detector is sampled, quantized and judged in each calculation time slot, and is subjected to serial-to-parallel conversionAnd restored to two-dimensional data. The calculation rate is 1/18 × 3 × 3 × 32 × 32/(32 × 32) ═ 5 × 108And MAC/s. By changing the length of the dispersion compensation fiber, the constraint relation among the modulation rate, the wavelength interval and the fiber length is destroyed, and the fault tolerance error of the system can be checked.
The foregoing embodiments may be modified in many different ways by those skilled in the art without departing from the spirit and scope of the invention, which is defined by the appended claims and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.
Claims (10)
1. A method for realizing multiplication and addition of an all-optical matrix based on multi-wavelength modulation and dispersion time delay is characterized in that a wavelength division multiplexing signal is used as a light source, modulated signals are weighted in a wavelength domain through a cascaded wavelength selective modulator after broadband modulation, or modulated signals are weighted in a time domain through a cascaded broadband modulator, so that multiplication operation is realized; and then, the weighted optical signals pass through a dispersion medium, so that dispersion delay superposition is carried out on each wavelength, and the addition operation of the signals is realized.
2. The method for implementing the all-optical matrix multiply-add based on the multi-wavelength modulation and the chromatic dispersion delay of claim 1, wherein the all-optical matrix multiply-add comprises: the wavelength domain weighting realizes 7-bit m-sequence autocorrelation operation, and the time domain weighting assisted by time division multiplexing realizes 4-order matrix vector multiplication and addition and 32 x 32 binary handwriting image handwriting edge extraction.
3. An integrated system on chip implementing the method of claim 1 or 2, comprising: light source, on-chip weighting ware, dispersion medium, photoelectric detector and the waveform generator that links to each other in proper order, wherein: each wavelength of the wavelength division multiplexing light source carries a signal to be weighted after broadband modulation, multiplication is realized through wavelength domain or time domain weighting, then corresponding time delay is carried out on each weighted wavelength signal by a designed dispersion medium, and multiplication is realized on a photoelectric detector.
4. The integrated system on a chip of claim 3, comprising: the device comprises a multi-wavelength optical frequency comb, a waveform shaper, a Mach-Zehnder modulator, a dispersion compensation optical fiber, a photoelectric detector and a pulse waveform generator which are connected in sequence.
5. The integrated system on a chip of claim 3, comprising: the optical fiber dispersion compensation device comprises a four-channel multi-wavelength light source, two cascaded Mach-Zehnder modulators, a dispersion compensation optical fiber, a photoelectric detector, an arbitrary waveform generator, an electric amplifier and an oscilloscope which are sequentially connected.
6. The integrated system on a chip of claim 3, comprising: the optical fiber dispersion compensation device comprises multiple wavelengths, two cascaded Mach-Zehnder modulators, a dispersion compensation optical fiber, a photoelectric detector, an arbitrary waveform generator, an electric amplifier and an oscilloscope which are sequentially connected.
7. The integrated system on a chip of claim 5, wherein the multi-wavelength light source is a non-equidistant optical frequency comb with respective intervals of 36GHz and 108 GHz; the dispersion coefficient of the dispersion compensation optical fiber is-160 ps/nm/km, and the length of the dispersion compensation optical fiber is 1.29 km; the pulse waveform generator generates a 7-bit m sequence of 16.875Gb/s, and a 6-bit isolation code element is additionally arranged to prevent the time slot overlapping coverage after dispersion delay; the modulation rate f, the wavelength interval delta lambda, the dispersion coefficient D and the optical fiber length l satisfy the following conditions: 1/f is D × Δ λ × l.
8. The system of claim 5, wherein the wavelengths of the multi-wavelength light source are 1550.12nm, 1553.12nm, 1556.12nm, 1559.12 nm; the arbitrary waveform generator generates a 4 × 4 matrix H and a 4 × 1 vector x of 5.37Gb/s in a time division multiplexing mode; the dispersion coefficient of the dispersion compensation fiber is-136 ps/nm/km, and the length of the dispersion compensation fiber is 455 m.
9. The integrated system on chip of claim 6, wherein the light source is 9 1550nm band wavelengths separated by 100GHz, the two-dimensional image and 3 x 3 laplacian are converted from parallel to serial into a one-dimensional data stream, wherein the laplacian is represented in 5-bit binary form and generates two cascaded mach-zehnder modulators modulating at 1Gb/s and 5Gb/s respectively in a time-division multiplexing manner by means of an arbitrary waveform generator; the dispersion coefficient of the dispersion compensation fiber is set to-160 ps/nm/km, and the length can be calculated to 7.8125 km; and the calculation result received by the photoelectric detector is subjected to sampling quantization and judgment in each calculation time slot, and is converted into two-dimensional data through serial-parallel conversion.
10. An application based on the integrated system on chip as claimed in any one of claims 3 to 9, characterized in that it is used as an arithmetic unit of a photonic neural network to form a feedforward neural network structure by a plurality of cascades.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910154200.5A CN109981172B (en) | 2019-03-01 | 2019-03-01 | All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910154200.5A CN109981172B (en) | 2019-03-01 | 2019-03-01 | All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109981172A CN109981172A (en) | 2019-07-05 |
CN109981172B true CN109981172B (en) | 2021-06-15 |
Family
ID=67077684
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910154200.5A Active CN109981172B (en) | 2019-03-01 | 2019-03-01 | All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109981172B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110838880B (en) * | 2019-11-12 | 2023-05-09 | 东南大学 | Efficient parallel broad-spectrum photon computing system and computing method |
CN113325917A (en) * | 2020-02-28 | 2021-08-31 | 华为技术有限公司 | Light computing device, system and computing method |
CN112001487A (en) * | 2020-07-20 | 2020-11-27 | 联合微电子中心有限责任公司 | Photon neural network |
CN112988113B (en) * | 2021-04-29 | 2021-09-14 | 中国科学院西安光学精密机械研究所 | Photon matrix vector multiplier |
CN114358271B (en) * | 2022-03-18 | 2022-07-12 | 之江实验室 | Time-wavelength interweaving photon neural network convolution acceleration chip |
CN115130666B (en) * | 2022-08-31 | 2022-11-22 | 之江实验室 | Two-dimensional photon convolution acceleration method and system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107870397A (en) * | 2016-09-26 | 2018-04-03 | 华为技术有限公司 | Wavelength selective optical switch |
CN108665924A (en) * | 2018-05-09 | 2018-10-16 | 上海交通大学 | Array silicon substrate programmable optical storage chip |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9564927B2 (en) * | 2015-05-27 | 2017-02-07 | John P Fonseka | Constrained interleaving for 5G wireless and optical transport networks |
-
2019
- 2019-03-01 CN CN201910154200.5A patent/CN109981172B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107870397A (en) * | 2016-09-26 | 2018-04-03 | 华为技术有限公司 | Wavelength selective optical switch |
CN108665924A (en) * | 2018-05-09 | 2018-10-16 | 上海交通大学 | Array silicon substrate programmable optical storage chip |
Also Published As
Publication number | Publication date |
---|---|
CN109981172A (en) | 2019-07-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109981172B (en) | All-optical matrix multiply-add implementation method based on multi-wavelength modulation and dispersion time delay | |
Amiri et al. | Nonlinear effects with semiconductor optical amplifiers | |
CN103678258B (en) | Method for improving data resolution ratio of silica-based optical matrix processor | |
TWI767877B (en) | Optoelectronic processing system | |
Shi et al. | InP photonic integrated multi-layer neural networks: Architecture and performance analysis | |
CN114815959B (en) | Photon tensor calculation acceleration method and device based on wavelength division multiplexing | |
Dai et al. | High channel-count comb filter based on chirped sampled fiber Bragg grating and phase shift | |
Leaird et al. | Generation of high-repetition-rate WDM pulse trains from an arrayed-waveguide grating | |
CN114819132B (en) | Photon two-dimensional convolution acceleration method and system based on time-wavelength interleaving | |
JP4733745B2 (en) | Optical signal processor | |
KR20140092687A (en) | Apparatus and method for controlling multi carrier light source generator | |
Jiang et al. | Photonic convolution neural network based on interleaved time-wavelength modulation | |
TW202215118A (en) | Optoelectronic processing apparatus, system and method | |
CN115169542A (en) | Two-dimensional photon convolution acceleration system and device for convolution neural network | |
Leaird et al. | Generation of flat-topped 500-GHz pulse bursts using loss engineered arrayed waveguide gratings | |
Sorokina | Multi-channel optical neuromorphic processor for frequency-multiplexed signals | |
Zhang et al. | High-speed parallel processing with photonic feedforward reservoir computing | |
CN115130666A (en) | Two-dimensional photon convolution acceleration method and system | |
Almeida et al. | All-optical packet compression based on time-to-wavelength conversion | |
Yang et al. | High-ratio electro-optical data compression for massive accessing networks using AOM-based ultrafast pulse shaping | |
JP2005070610A (en) | Multi-wavelength light source apparatus | |
Mourgias-Alexandris et al. | Experimental demonstration of an optical neuron with a logistic Sigmoid activation function | |
Bajaj et al. | Application of soft computing technique for optical add/drop multiplexer in terms of OSNR and optical noise power | |
Shinya et al. | High-Speed Optical Convolutional Neural Network Accelerator with 100 Gbaud EO-polymer/Si Hybrid Optical Modulator | |
Nakajima et al. | Densely Parallelized Photonic Tensor Processor on Hybrid Waveguide/Free-Space-Optics |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |