A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network

Abstract

The application provides a matrix operation accelerator combining wavelength division multiplexing and MZI cascade network, relates to the field of optical neural networks, and comprises an input signal layer, a weight signal layer, a summation layer and a nonlinear layer.The input signal layer is used for realizing matrix operation of optical signals through a Mach-Zehnder interferometer array; the weight signal layer is used for applying an electrical signal to a micro-ring modulator array to adjust a weight signal; the summation layer is used for separating the results of the action of the optical signals of different wavelengths through the weight signal; and the nonlinear layer is used for converting the optical signals into electrical signals through a photodetector array to realize a nonlinear activation function in the electrical domain.The application introduces N different wavelengths in the network formed by MZI cascade, so that the number of times of executing matrix operation is increased by N times each time, high-speed convolution operation is facilitated, and the size of the micro-ring modulator is relatively small, so that the energy efficiency and area ratio of the MZI cascade network calculation can be effectively increased.

Description

Technical Field

[0001] This application relates to the field of optical neural networks, and more particularly to a matrix operation accelerator that combines wavelength division multiplexing and MZI cascaded networks. Background Technology

[0002] With the rapid development of information technology, massive amounts of data and information are constantly being generated, providing a solid foundation for the rapid development of neural networks. As a result, neural networks have achieved remarkable success in computer vision, natural language processing, and medical imaging. However, processing and computing this massive amount of information places higher demands on hardware performance, requiring higher computing speeds and lower power consumption. But as the feature size of transistors gradually approaches the physical limit and enters the nanometer scale, the development of integrated circuits has encountered a bottleneck, making it difficult to continue Moore's Law.

[0003] Because photonic devices offer higher bandwidth and lower power consumption compared to electronic devices, they are being used to implement matrix operations in neural networks. Currently, the main methods for implementing optical neural networks include cascaded Mach-Zehnder arrays and micro-ring modulator arrays.

[0004] However, the number of devices required for matrix operations based on Mach-Zehnder interferometers is proportional to the square of the matrix dimension, and the devices are also relatively large, making it difficult to achieve large-scale matrix operations. Furthermore, when performing non-square matrix operations (i.e., when the number of rows is not equal to the number of columns), redundant Mach-Zehnder interferometers are used, reducing the efficiency of performing effective matrix operations. Summary of the Invention

[0005] To address the above problems, a matrix operation accelerator combining wavelength division multiplexing (WDM) and MZI cascaded networks is proposed, including:

[0006] The input signal layer is used to perform matrix operations on optical signals using a Mach-Zehnder interferometer array.

[0007] The weighted signal layer is used to apply electrical signals to the micro-ring modulator array to adjust the weighted signals;

[0008] The summation layer is used to separate the results of the weighting signal on optical signals of different wavelengths.

[0009] The nonlinear layer is used to convert the optical signal into an electrical signal through a photodetector array, and to implement a nonlinear activation function in the electrical domain.

[0010] Optionally, the input signal layer includes:

[0011] A semiconductor laser, used to emit optical signals of different wavelengths;

[0012] A wavelength division multiplexing device is used to combine optical signals of different wavelengths in the same waveguide;

[0013] A multimode interferometer module is used to distribute the optical signal evenly to each input port of the Mach-Zehnder interferometer array according to power.

[0014] A Mach-Zehnder interferometer array is used to perform matrix operations on the combined optical signals.

[0015] Optionally, the semiconductor laser includes:

[0016] Distributed feedback semiconductor laser;

[0017] Distributed Bragg reflector semiconductor laser;

[0018] Vertical cavity surface-emitting laser.

[0019] Optionally, the wavelength division multiplexing device is an arrayed waveguide grating or a cascaded Mach-Zehnder interferometer type multiplexer.

[0020] Optionally, the multimode interferometer module can be a 1×N multimode interferometer or a cascaded log2N 1×2 multimode interferometers.

[0021] Optionally, the Mach-Zehnder interferometer array includes a unitary matrix array 1, a diagonal matrix array 2, and a unitary matrix array 3, wherein,

[0022] The unitary matrix array 1 includes N(N-1) / 2 cascaded arrays of first Mach-Zehnder interferometers;

[0023] The diagonal matrix array 2 includes N cascaded arrays of second Mach-Zehnder interferometers;

[0024] The unitary matrix array 3 includes N(N-1) / 2 cascaded third Mach-Zehnder interferometer arrays;

[0025] Where N is the dimension of the matrix to be operated on.

[0026] Optionally, the unitary matrix array 1, the diagonal matrix array 2, and the unitary matrix array 3 are composed of multiple identical Mach-Zehnder interferometers, wherein,

[0027] The Mach-Zehnder interferometer includes a coupler module and a phase shifter module;

[0028] The coupler module is composed of a multimode interferometer or a directional coupler, and the phase shifter module realizes its function based on the thermo-optic effect or the electro-optic effect.

[0029] Optionally, the weighted signal layer includes the micro-ring modulator array and the cross-waveguide array, wherein,

[0030] The micro-ring modulator array includes multiple micro-ring modulators, and the cross waveguide array includes multiple cross waveguides;

[0031] The micro-ring modulator and the cross waveguide constitute multiple basic components. The optical signal is input through the in1 or in2 port of the basic component, and the optical signal is output along the out1 or out2 port by adjusting the micro-ring modulator.

[0032] Optionally, the photodetector array includes:

[0033] A first subarray of photodetectors is used to calibrate the resonant wavelength of each of the microring modulators;

[0034] The second subarray of the photodetector is used to detect the power value of each output port and implements a nonlinear activation function in the electric domain.

[0035] The technical solutions provided by the embodiments of this application bring at least the following beneficial effects:

[0036] By introducing N different wavelengths into the network formed by MZI cascades, the number of matrix operations performed each time is increased by N times, which is beneficial for performing high-speed convolution operations. Furthermore, since the size of the micro-ring modulator is relatively small, it can effectively increase the energy efficiency and area ratio of the MZI cascade network computation.

[0037] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0038] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0039] Figure 1 This is a structural diagram of a matrix operation accelerator combining wavelength division multiplexing and MZI cascaded networks, according to an embodiment of this application.

[0040] Figure 2 This is a structural diagram of a Mach-Zehnder interferometer according to an embodiment of this application;

[0041] Figure 3 This is a structural diagram of the basic constituent units shown according to an embodiment of this application. Detailed Implementation

[0042] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0043] Figure 1 This is a structural diagram of a matrix operation accelerator combining wavelength division multiplexing and an MZI cascaded network, according to an embodiment of this application, comprising:

[0044] The input signal layer is used to perform matrix operations on optical signals using a Mach-Zehnder interferometer array.

[0045] The weighted signal layer is used to apply electrical signals to the micro-ring modulator array to adjust the weighted signals;

[0046] The summation layer is used to separate the results of light signals of different wavelengths after being processed by the weighting signal.

[0047] Nonlinear layers are used to convert optical signals into electrical signals through a photodetector array, and to implement nonlinear activation functions in the electric domain.

[0048] The structure and function of each component are described in detail below.

[0049] like Figure 1 As shown, the input signal layer includes semiconductor lasers LD1-LD3, wavelength division multiplexing device Demux1, multimode interferometer module MMI, and Mach-Zehnder interferometer arrays MZI1-MZI9.

[0050] Specifically, LD1-LD3 are three semiconductor lasers with different emission wavelengths. These semiconductor lasers include distributed feedback semiconductor lasers, distributed Bragg reflector semiconductor lasers, and vertical cavity surface-emitting lasers.

[0051] In one possible embodiment, LD1 is a distributed Bragg reflector semiconductor laser.

[0052] Demux1 is a wavelength division multiplexing device that combines optical signals of different wavelengths in the same waveguide. Wavelength division multiplexing devices include arrayed waveguide gratings or cascaded Mach-Zehnder interferometer-type multiplexers.

[0053] In one possible embodiment, Demux1 is an arrayed waveguide grating.

[0054] MMI utilizes its wavelength insensitivity to distribute the combined optical signal evenly to each output port according to power. The multimode interferometer module is a 1×N multimode interferometer or composed of log2N cascaded 1×2 multimode interferometers.

[0055] In one possible embodiment, the MMI is a 1×N multimode interferometer.

[0056] In addition, the Mach-Zehnder interferometer array includes unitary matrix array 1, diagonal matrix array 2, and unitary matrix array 3, among which,

[0057] The unitary matrix array 1 comprises N(N-1) / 2 cascaded arrays of first Mach-Zehnder interferometers;

[0058] The diagonal matrix array 2 comprises N cascaded arrays of second Mach-Zehnder interferometers;

[0059] The unitary matrix array 3 includes N(N-1) / 2 cascaded third Mach-Zehnder interferometer arrays;

[0060] Where N is the dimension of the matrix to be operated on.

[0061] like Figure 2 As shown, MZI1-MZI9 are Mach-Zehnder interferometers used to perform operations on arbitrary matrices. Among them, MZI1-MZI3 are unitary matrix array 1, MZI4-MZI6 are diagonal matrix array 2, and MZI7-MZI9 are unitary matrix array 3. MZI1-MZI3 implement the function of unitary matrix 1, MZI4-MZI6 implement the function of diagonal matrix 2, and MZI7-MZI9 implement the function of unitary matrix 3.

[0062] In addition to using a Mach-Zehnder interferometer, MZI4-MZI6 can also be achieved using a tunable optical attenuator.

[0063] In the embodiments of this application, each MZI is composed of multiple identical Mach-Zehnder interferometers, such as Figure 2 Its components include two 2×2 couplers, coupler1 and coupler2, and two phase shifters, ps1 and ps2. The couplers are implemented by a multimode interferometer or a directional coupler, and the phase shifters are implemented by thermo-optical effect or electro-optical effect.

[0064] like Figure 1 As shown, MRM1-MRM9 are micro-ring modulator arrays, and Cross1-Cross9 are cross-waveguide arrays. The basic components of the micro-ring modulators and cross-waveguides are as follows: Figure 3 As shown, the optical signal is input through port in1 or in2, and the optical signal is output along ports out1 and out2 by adjusting the signal of the micro-ring modulator.

[0065] In the N×N microring modulator array, the resonant wavelengths of the microring modulators in each column are the same, with the resonant wavelengths from the first column to the Nth column being λ1, λ2, ..., λ3 respectively. N .

[0066] like Figure 1 As shown, PD1-PD6 are photodetector arrays, where PD1-PD3 is the first subarray of photodetectors used to calibrate the resonant wavelength of each micro-ring modulator, and PD4-PD6 is the second subarray of photodetectors used to detect the power value of each output port and implement a nonlinear activation function in the electric domain.

[0067] The embodiments of this application introduce N different wavelengths into the network formed by MZI cascade, which increases the number of matrix operations performed each time by N times. This is beneficial for performing high-speed convolution operations. Furthermore, since the size of the micro-ring modulator is relatively small, it can effectively increase the energy efficiency and area ratio of the MZI cascade network computation.

[0068] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0069] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A matrix operation accelerator combining wavelength division multiplexing and MZI cascaded network, characterized in that, The application relates to a photonic neural network chip, which comprises the following layers: an input signal layer, a weight signal layer, a summation layer and a nonlinear layer. The input signal layer is used for realizing matrix operation of optical signals through a Mach-Zehnder interferometer array. The weight signal layer is used for applying electrical signals to a micro-ring modulator array to adjust weight signals. The summation layer is used for separating the results of the optical signals of different wavelengths through the weight signals. The nonlinear layer is used for converting the optical signals into electrical signals through a photodetector array to realize a nonlinear activation function in the electrical domain. The input signal layer comprises the following components: A semiconductor laser is used for emitting optical signals of different wavelengths. A wavelength division multiplexing device is used for combining the optical signals of different wavelengths in the same waveguide. A multimode interferometer module is used for equally distributing the optical signals to each input port of the Mach-Zehnder interferometer array according to power. The Mach-Zehnder interferometer array realizes matrix operation of the combined optical signals. The Mach-Zehnder interferometer array comprises a unitary matrix array 1, a diagonal matrix array 2 and a unitary matrix array 3. The unitary matrix array 1 comprises N (N-1) / 2 first Mach-Zehnder interferometer arrays in cascade. The diagonal matrix array 2 comprises N second Mach-Zehnder interferometer arrays in cascade. The unitary matrix array 3 comprises N (N-1) / 2 third Mach-Zehnder interferometer arrays in cascade. N is the dimension of the matrix to be operated. The weight signal layer comprises the micro-ring modulator array and a cross waveguide array. The micro-ring modulator array comprises a plurality of micro-ring modulators, and the cross waveguide array comprises a plurality of cross waveguides. The micro-ring modulator and the cross waveguide constitute a plurality of basic units, the optical signals are input from the in1 or in2 port of the basic unit, and the optical signals are output along the out1 or out2 port through the adjustment of the micro-ring modulator. The photodetector array comprises: A photodetector first subarray is used for calibrating the resonant wavelength of each micro-ring modulator. A photodetector second subarray is used for detecting the power value of each output port to realize a nonlinear activation function in the electrical domain.

2. The matrix operation accelerator combined with a wavelength division multiplexing and MZI cascade network according to claim 1, characterized in that, The semiconductor laser comprises: A distributed feedback semiconductor laser, A distributed Bragg reflection semiconductor laser, A vertical cavity surface emitting laser.

3. The matrix operation accelerator combined with a wavelength division multiplexing and MZI cascade network according to claim 1, characterized in that, The wavelength division multiplexing device is an arrayed waveguide grating or a multiplexer of cascaded Mach-Zehnder interferometers.

4. The matrix operation accelerator combined with a wavelength division multiplexing and MZI cascade network according to claim 1, characterized in that, The multimode interferometer module is a 1xN multimode interferometer or is composed of log2N 1x2 multimode interferometers in cascade.

5. The matrix operation accelerator coupled with a WDM and MZI cascaded network according to claim 1, wherein, The unitary matrix array 1, the diagonal matrix array 2 and the unitary matrix array 3 are composed of a plurality of identical Mach-Zehnder interferometers. The Mach-Zehnder interferometer comprises a coupler module and a phase shifter module. The coupler module is composed of a multimode interferometer or a directional coupler, and the phase shifter module realizes functions according to the thermo-optic effect or the electro-optic effect.

Images