CN116822601A - A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network - Google Patents
A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network Download PDFInfo
- Publication number
- CN116822601A CN116822601A CN202310642207.8A CN202310642207A CN116822601A CN 116822601 A CN116822601 A CN 116822601A CN 202310642207 A CN202310642207 A CN 202310642207A CN 116822601 A CN116822601 A CN 116822601A
- Authority
- CN
- China
- Prior art keywords
- array
- division multiplexing
- mach
- wavelength division
- matrix operation
- 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
Landscapes
- Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
Abstract
Description
技术领域Technical field
本申请涉及光学神经网络领域,尤其涉及一种结合波分复用和MZI级联网络的矩阵运算加速器。This application relates to the field of optical neural networks, and in particular to a matrix operation accelerator that combines wavelength division multiplexing and MZI cascade networks.
背景技术Background technique
随着信息技术的快速发展,海量的数据和信息在不断的产生,这为神经网络的迅猛发展提供良好的基础。因此神经网络在计算机视觉,自然语言处理以及医学成像等方面都取得了不菲的成就。然而面对这些海量信息的处理和计算都对硬件的性能提出了更高的要求,需要其具备更高的计算速度和更低的功耗。但随着晶体管的特征尺寸逐渐逼近物理极限步入纳米量级,集成电路的发展也遇到了瓶颈,难以延续摩尔定律。With the rapid development of information technology, massive amounts of data and information are constantly being generated, which provides a good foundation for the rapid development of neural networks. Therefore, neural networks have made considerable achievements in computer vision, natural language processing, and medical imaging. However, the processing and calculation of these massive amounts of information place higher requirements on hardware performance, requiring higher computing speed and lower power consumption. However, as the characteristic size of transistors gradually approaches the physical limit and enters the nanometer level, the development of integrated circuits has also encountered a bottleneck, making it difficult to continue Moore's Law.
由于光子器件相较于电子器件具有更高带宽和更低能耗,因此光子器件也正在被用于实现神经网络中矩阵运算的功能。目前实现光神经网络的主要包括级联的马赫曾德尔阵列以及微环调制器阵列。Since photonic devices have higher bandwidth and lower energy consumption than electronic devices, photonic devices are also being used to implement matrix operations in neural networks. The current implementation of optical neural networks mainly includes cascaded Mach-Zehnder arrays and micro-ring modulator arrays.
然而基于马赫曾德尔干涉仪实现矩阵运算的方案所需要的器件数量与矩阵的维度成平方关系,而且器件的尺寸也较大,因此难以实现较大规模的矩阵运算功能。在实现非方阵运算时即矩阵的行数不等于列数时还会有冗余的马赫曾德尔干涉仪,降低执行有效矩阵运算的效率。However, the number of devices required to implement matrix operations based on Mach-Zehnder interferometers is squarely related to the dimensions of the matrix, and the size of the devices is also large, so it is difficult to implement large-scale matrix operations. When implementing non-square matrix operations, that is, when the number of rows of the matrix is not equal to the number of columns, there will be redundant Mach-Zehnder interferometers, which reduces the efficiency of performing effective matrix operations.
发明内容Contents of the invention
针对上述问题,提出了一种结合波分复用和MZI级联网络的矩阵运算加速器,包括:In response to the above problems, a matrix operation accelerator combining wavelength division multiplexing and MZI cascade network is proposed, including:
输入信号层,用于通过马赫曾德尔干涉仪阵列实现光信号的矩阵运算;The input signal layer is used to implement matrix operations of optical signals through the Mach-Zehnder interferometer array;
权重信号层,用于为微环调制器阵列施加电信号,以调节权重信号;The weight signal layer is used to apply electrical signals to the microring modulator array to adjust the weight signal;
求和层,用于将不同波长的所述光信号经过所述权重信号的作用结果分离开;A summation layer, used to separate the optical signals of different wavelengths through the action results of the weighting signal;
非线性层,用于通过光电探测器阵列将所述光信号转换为电信号,在电域内实现非线性激活函数。The nonlinear layer is used to convert the optical signal into an electrical signal through the photodetector array, and implement a nonlinear activation function in the electrical domain.
可选的,所述输入信号层,包括:Optionally, the input signal layer includes:
半导体激光器,用于发射不同波长的所述光信号;Semiconductor laser, used to emit the optical signals of different wavelengths;
波分复用器件,用于将所述不同波长的所述光信号合波在同一波导中;A wavelength division multiplexing device for multiplexing the optical signals of different wavelengths in the same waveguide;
多模干涉仪模块,用于将所述光信号按照功率平均分配至所述马赫曾德尔干涉仪阵列的各个输入端口;A multi-mode interferometer module, configured to evenly distribute the optical signal to each input port of the Mach-Zehnder interferometer array according to power;
马赫曾德尔干涉仪阵列,实现所述合波光信号的矩阵运算。The Mach-Zehnder interferometer array realizes the matrix operation of the multiplexed optical signal.
可选的,所述半导体激光器,包括:Optionally, the semiconductor laser includes:
分布反馈式半导体激光器;Distributed feedback semiconductor laser;
分布式布拉格反射半导体激光器;Distributed Bragg reflection semiconductor laser;
垂直腔面发射激光器。Vertical cavity surface emitting laser.
可选的,所述波分复用器件为阵列波导光栅或者级联的马赫曾德尔干涉仪型复用器。Optionally, the wavelength division multiplexing device is an arrayed waveguide grating or a cascaded Mach-Zehnder interferometer type multiplexer.
可选的,所述多模干涉仪模块为一个1×N的多模干涉仪或由log2N个1×2的多模干涉仪级联组成。Optionally, the multi-mode interferometer module is a 1×N multi-mode interferometer or is composed of a cascade of log 2 N 1×2 multi-mode interferometers.
可选的,所述马赫曾德尔干涉仪阵列包括酉矩阵阵列1、对角矩阵阵列2和酉矩阵阵列3,其中,Optionally, the Mach-Zehnder interferometer array includes unitary matrix array 1, diagonal matrix array 2 and unitary matrix array 3, where,
所述酉矩阵阵列1包括级联的N(N-1)/2个的第一马赫曾德尔干涉仪阵列;The unitary matrix array 1 includes a cascade of N(N-1)/2 first Mach-Zehnder interferometer arrays;
所述对角矩阵阵列2包括级联的N个第二马赫曾德尔干涉仪阵列;The diagonal matrix array 2 includes a cascade of N second Mach-Zehnder interferometer arrays;
所述酉矩阵阵列3包括级联的N(N-1)/2个的第三马赫曾德尔干涉仪阵列;The unitary matrix array 3 includes a cascade of N(N-1)/2 third Mach-Zehnder interferometer arrays;
其中,N为待运算矩阵的维数。Among them, N is the dimension of the matrix to be operated.
可选的,所述酉矩阵阵列1、对角矩阵阵列2和酉矩阵阵列3由多个相同的马赫曾德尔干涉仪组成,其中,Optionally, the unitary matrix array 1, diagonal matrix array 2 and unitary matrix array 3 are composed of multiple identical Mach-Zehnder interferometers, wherein,
所述马赫曾德尔干涉仪包括耦合器模块与移相器模块;The Mach-Zehnder interferometer includes a coupler module and a phase shifter module;
所述耦合器模块由多模干涉仪或定向耦合器构成,所述移相器模块根据热光效应或电光效应实现功能。The coupler module is composed of a multi-mode interferometer or a directional coupler, and the phase shifter module implements functions based on thermo-optical effects or electro-optical effects.
可选的,所述权重信号层包括所述微环调制器阵列与交叉波导阵列,其中,Optionally, the weighted signal layer includes the micro-ring modulator array and the crossed waveguide array, wherein,
所述微环调制器阵列包括多个微环调制器,所述交叉波导阵列包括多个交叉波导;The microring modulator array includes a plurality of microring modulators, and the crossed waveguide array includes a plurality of crossed waveguides;
所述微环调制器与所述交叉波导构成多个基本组成单元,其中,所述光信号由所述基本组成单元的in1或in2端口输入,通过调节所述微环调制器使得所述光信号沿着out1或out2端口输出。The micro-ring modulator and the cross waveguide form a plurality of basic components, wherein the optical signal is input from the in1 or in2 port of the basic component unit, and the optical signal is adjusted by adjusting the micro-ring modulator. Output along the out1 or out2 port.
可选的,所述光电探测器阵列,包括:Optionally, the photodetector array includes:
光电探测器第一子阵列,用于校准每个所述微环调制器的谐振波长;a first sub-array of photodetectors, used to calibrate the resonant wavelength of each of the micro-ring modulators;
光电探测器第二子阵列,用于探测每个输出端口的功率值,在电域内实现非线性激活函数。The second sub-array of photodetectors is used to detect the power value of each output port and implement a nonlinear activation function in the electrical domain.
本申请的实施例提供的技术方案至少带来以下有益效果:The technical solutions provided by the embodiments of the present application at least bring the following beneficial effects:
通过在MZI级联形成的网络中通过引入N个不同的波长使得每次实现执行矩阵运算的次数扩大N倍,有利于执行高速的卷积运算,且由于微环调制器的尺寸相对较小,能有效增加MZI级联网络计算的能效与面积比。By introducing N different wavelengths into the network formed by the MZI cascade, the number of matrix operations performed each time is expanded by N times, which is conducive to performing high-speed convolution operations, and due to the relatively small size of the microring modulator, It can effectively increase the energy efficiency and area ratio of MZI cascade network computing.
本申请附加的方面和优点将在下面的描述中部分给出,部分将从下面的描述中变得明显,或通过本申请的实践了解到。Additional aspects and advantages of the 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 the application.
附图说明Description of the drawings
本申请上述的和/或附加的方面和优点从下面结合附图对实施例的描述中将变得明显和容易理解,其中:The above and/or additional aspects and advantages of the present application will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:
图1是根据本申请实施例示出的一种结合波分复用和MZI级联网络的矩阵运算加速器的结构图;Figure 1 is a structural diagram of a matrix operation accelerator that combines wavelength division multiplexing and MZI cascade networks according to an embodiment of the present application;
图2是根据本申请实施例示出的马赫曾德尔干涉仪的结构图;Figure 2 is a structural diagram of a Mach-Zehnder interferometer according to an embodiment of the present application;
图3是根据本申请实施例示出的基本组成单元的结构图。Figure 3 is a structural diagram of a basic component unit according to an embodiment of the present application.
具体实施方式Detailed ways
下面详细描述本申请的实施例,实施例的示例在附图中示出,其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的,旨在用于解释本申请,而不能理解为对本申请的限制。The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present application, but should not be construed as limiting the present application.
图1是根据本申请实施例示出的一种结合波分复用和MZI级联网络的矩阵运算加速器的结构图,包括:Figure 1 is a structural diagram of a matrix operation accelerator that combines wavelength division multiplexing and MZI cascade networks according to an embodiment of the present application, including:
输入信号层,用于通过马赫曾德尔干涉仪阵列实现光信号的矩阵运算;The input signal layer is used to implement matrix operations of optical signals through the Mach-Zehnder interferometer array;
权重信号层,用于为微环调制器阵列施加电信号,以调节权重信号;The weight signal layer is used to apply electrical signals to the microring modulator array to adjust the weight signal;
求和层,用于将不同波长的光信号经过权重信号的作用结果分离开;The summation layer is used to separate the effects of weighted signals on optical signals of different wavelengths;
非线性层,用于通过光电探测器阵列将光信号转换为电信号,在电域内实现非线性激活函数。The nonlinear layer is used to convert optical signals into electrical signals through the photodetector array and implement nonlinear activation functions in the electrical domain.
下面详细介绍各个部件的结构与作用。The structure and function of each component are introduced in detail below.
如图1所示,输入信号层包括半导体激光器LD1-LD3、波分复用器件Demux1、多模干涉仪模块MMI和马赫曾德尔干涉仪阵列MZI1-MZI9。As shown in Figure 1, the input signal layer includes semiconductor lasers LD1-LD3, wavelength division multiplexing device Demux1, multi-mode interferometer module MMI and Mach-Zehnder interferometer array MZI1-MZI9.
具体的说,LD1-LD3为三个不同发光波长的半导体激光器,半导体激光器包括分布反馈式半导体激光器,分布式布拉格反射半导体激光器以及垂直腔面发射激光器。Specifically, LD1-LD3 are three semiconductor lasers with different emission wavelengths. The semiconductor lasers include distributed feedback semiconductor lasers, distributed Bragg reflection semiconductor lasers and vertical cavity surface emitting lasers.
一种可能的实施例中,LD1为分布式布拉格反射半导体激光器。In a possible embodiment, LD1 is a distributed Bragg reflection semiconductor laser.
其中,Demux1为波分复用器件,将不同波长的光信号合波在同一波导中,波分复用器件包括阵列波导光栅或者级联的马赫曾德尔干涉仪型复用器。Among them, Demux1 is a wavelength division multiplexing device that multiplexes optical signals of different wavelengths in the same waveguide. The wavelength division multiplexing device includes an arrayed waveguide grating or a cascaded Mach-Zehnder interferometer type multiplexer.
一种可能的实施例中,Demux1为阵列波导光栅。In a possible embodiment, Demux1 is an arrayed waveguide grating.
MMI利用其波长不敏感特性将合波的光信号按照功率平均分配每个输出端口,多模干涉仪模块为一个1×N的多模干涉仪或由log2N个1×2的多模干涉仪级联组成。MMI uses its wavelength insensitivity feature to evenly distribute the combined optical signals to each output port according to power. The multimode interferometer module is a 1×N multimode interferometer or consists of log 2 N 1×2 multimode interferences. Instrument cascade composition.
一种可能的实施例中,MMI为一个1×N的多模干涉仪。In a possible embodiment, the MMI is a 1×N multi-mode interferometer.
另外,马赫曾德尔干涉仪阵列包括酉矩阵阵列1、对角矩阵阵列2和酉矩阵阵列3,其中,In addition, the Mach-Zehnder interferometer array includes unitary matrix array 1, diagonal matrix array 2 and unitary matrix array 3, where,
酉矩阵阵列1包括级联的N(N-1)/2个的第一马赫曾德尔干涉仪阵列;The unitary matrix array 1 includes a cascade of N(N-1)/2 first Mach-Zehnder interferometer arrays;
对角矩阵阵列2包括级联的N个第二马赫曾德尔干涉仪阵列;The diagonal matrix array 2 includes a cascade of N second Mach-Zehnder interferometer arrays;
酉矩阵阵列3包括级联的N(N-1)/2个的第三马赫曾德尔干涉仪阵列;The unitary matrix array 3 includes a cascaded N(N-1)/2 third Mach-Zehnder interferometer array;
其中,N为待运算矩阵的维数。Among them, N is the dimension of the matrix to be operated.
如图2所示,MZI1-MZI9为马赫曾德尔干涉仪,用于实现任意矩阵的运算,其中,MZI1-MZI3为酉矩阵阵列1,MZI4-MZI6为对角矩阵阵列2,MZI7-MZI9为酉矩阵阵列3,MZI1-MZI3实现的酉矩阵1的功能,MZI4-MZI6实现对角矩阵2的功能,MZI7-MZI9则是实现酉矩阵3的功能.As shown in Figure 2, MZI1-MZI9 are Mach-Zehnder interferometers, used to implement arbitrary matrix operations. Among them, MZI1-MZI3 are unitary matrix array 1, MZI4-MZI6 are diagonal matrix array 2, and MZI7-MZI9 are unitary matrix arrays. Matrix array 3, MZI1-MZI3 realize the function of unitary matrix 1, MZI4-MZI6 realize the function of diagonal matrix 2, MZI7-MZI9 realize the function of unitary matrix 3.
其中,MZI4-MZI6除了使用马赫曾德尔干涉仪之外还可以利用可调光衰减器实现。Among them, MZI4-MZI6 can be realized by using an adjustable optical attenuator in addition to using a Mach-Zehnder interferometer.
本申请实施例中,每个MZI都是多个相同的马赫曾德尔干涉仪组成,如图2,其组成部分包括coupler1和coupler2这两个2×2的耦合器,以及ps1和ps2这两个移相器,耦合器由多模干涉仪或者定向耦合器实现,移相器由热光效应或者电光效应实现功能。In the embodiment of this application, each MZI is composed of multiple identical Mach-Zehnder interferometers, as shown in Figure 2. Its components include two 2×2 couplers, coupler1 and coupler2, and two ps1 and ps2. Phase shifters and couplers are implemented by multi-mode interferometers or directional couplers, and phase shifters are implemented by thermo-optical effects or electro-optical effects.
如图1所示,MRM1-MRM9为微环调制器阵列,Cross1-Cross9为交叉波导阵列,其中,微环调制器与交叉波导的基本组成单元如图3所示,光信号由in1或者in2口输入,通过调节微环调制器的信号使得光信号沿着out1和out2端口输出。As shown in Figure 1, MRM1-MRM9 are micro-ring modulator arrays, and Cross1-Cross9 are crossed waveguide arrays. The basic components of the micro-ring modulator and cross waveguide are shown in Figure 3. The optical signal is transmitted through the in1 or in2 port. Input, by adjusting the signal of the microring modulator, the optical signal is output along the out1 and out2 ports.
其中,N×N的微环调制器阵列中,每一列的微环调制器的谐振波长都相同,从第一列到第N列的谐振波长分别为λ1,λ2……λN。Among them, in the N×N micro-ring modulator array, the resonant wavelengths of the micro-ring modulators in each column are the same, and the resonant wavelengths from the first column to the N-th column are λ 1 , λ 2 ...λ N respectively.
如图1所示,PD1-PD6为光电探测器阵列,其中,PD1-PD3为光电探测器第一子阵列,用于校准每个微环调制器的谐振波长,PD4-PD6为光电探测器第二子阵列,用于探测每个输出端口的功率值,并且在电域内实现非线性激活函数。As shown in Figure 1, PD1-PD6 are photodetector arrays, in which PD1-PD3 are the first sub-array of photodetectors, used to calibrate the resonant wavelength of each micro-ring modulator, and PD4-PD6 are the first sub-arrays of photodetectors. Two sub-arrays are used to detect the power value of each output port and implement a nonlinear activation function in the electrical domain.
本申请实施例通过在MZI级联形成的网络中通过引入N个不同的波长使得每次实现执行矩阵运算的次数扩大N倍,有利于执行高速的卷积运算,且由于微环调制器的尺寸相对较小,能有效增加MZI级联网络计算的能效与面积比。The embodiment of the present application expands the number of matrix operations by N times by introducing N different wavelengths into the network formed by the MZI cascade, which is conducive to performing high-speed convolution operations, and due to the size of the micro-ring modulator Relatively small, it can effectively increase the energy efficiency and area ratio of MZI cascade network computing.
应该理解,可以使用上面所示的各种形式的流程,重新排序、增加或删除步骤。例如,本发公开中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行,只要能够实现本公开公开的技术方案所期望的结果,本文在此不进行限制。It should be understood that various forms of the process shown above may be used, with steps reordered, added or deleted. For example, each step described in the present disclosure can be executed in parallel, sequentially, or in a different order. As long as the desired results of the technical solution disclosed in the present disclosure can be achieved, there is no limitation here.
上述具体实施方式,并不构成对本公开保护范围的限制。本领域技术人员应该明白的是,根据设计要求和其他因素,可以进行各种修改、组合、子组合和替代。任何在本公开的精神和原则之内所作的修改、等同替换和改进等,均应包含在本公开保护范围之内。The above-mentioned specific embodiments do not constitute a limitation on the scope of the present disclosure. It will be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions are possible depending on design requirements and other factors. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this disclosure shall be included in the protection scope of this disclosure.
Claims (9)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310642207.8A CN116822601A (en) | 2023-05-31 | 2023-05-31 | A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310642207.8A CN116822601A (en) | 2023-05-31 | 2023-05-31 | A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116822601A true CN116822601A (en) | 2023-09-29 |
Family
ID=88121320
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310642207.8A Pending CN116822601A (en) | 2023-05-31 | 2023-05-31 | A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116822601A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN118244555A (en) * | 2024-05-27 | 2024-06-25 | 武汉大学 | All-optical nonlinear activator, implementation method thereof, network structure and computing chip |
-
2023
- 2023-05-31 CN CN202310642207.8A patent/CN116822601A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN118244555A (en) * | 2024-05-27 | 2024-06-25 | 武汉大学 | All-optical nonlinear activator, implementation method thereof, network structure and computing chip |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022016894A1 (en) | Photonic neural network | |
CN116348886B (en) | Coherent Photonic Computing Architecture | |
Bai et al. | Photonic multiplexing techniques for neuromorphic computing | |
TWI758994B (en) | Optoelectronic processing system | |
JP2024503991A (en) | Balanced photonic architecture for matrix computation | |
Peserico et al. | Integrated photonic tensor processing unit for a matrix multiply: a review | |
Totovic et al. | WDM equipped universal linear optics for programmable neuromorphic photonic processors | |
CN113805641B (en) | Photonic neural network | |
WO2023201970A1 (en) | Computing chip, system, and data processing method | |
Tsakyridis et al. | Universal linear optics for ultra-fast neuromorphic silicon photonics towards Fj/MAC and TMAC/sec/mm 2 engines | |
CN110908428A (en) | Parallel optical computing system for efficiently realizing large-scale matrix operation | |
CN116822601A (en) | A matrix operation accelerator combining wavelength division multiplexing and MZI cascade network | |
Soref | Reconfigurable integrated optoelectronics | |
CN115222035A (en) | A photonic neural network convolution acceleration chip | |
KR20230042333A (en) | Systems and methods for exploiting photonic degrees of freedom in a photonic processor | |
US20240013041A1 (en) | Single ended eam with electrical combining | |
CN116739063A (en) | A neural network accelerator based on multi-mode interferometer and coherent detection | |
CN111538368A (en) | Photon information processing chip | |
CN113392965B (en) | A method, device and storage medium for realizing Hadamard product | |
CN112799464B (en) | Matrix vector multiplier based on multi-wavelength light source and operation method thereof | |
Pappas et al. | Reaching the Peta-computing: 163.8 TOPS through Multidimensional AWGR-based Accelerators | |
De Marinis et al. | A silicon nitride reconfigurable linear optical processor | |
CN117255967A (en) | Balanced photon architecture for matrix computation | |
Ikeda et al. | High-fidelity WDM-compatible photonic processor for matrix-matrix multiplication | |
Kumar et al. | All optical 2 to 1 multiplexer using optical tree architecture |
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 |