Design Of Computing Hardware Support For Hybrid Optical-electronic Neural Networks Based On FPGA

In recent years,the study of artificial neural network and its application has become a hot topic all over the world.The brain consists of a huge and complex network of neurons.Nerve cells are collected by types to form layers to achieve advanced information processing.Neural networks are designed to artificially replicate the network of neurons that make up the human brain.Information processing and solving various problems,such as image recognition,traffic prediction,medical discovery,etc.Since neural networks often extract features by increasing the number of intermediate layers,the more and deeper the intermediate layers are,the more likely it is to distinguish and identify.Therefore,as the application of neural networks becomes more and more complex,the scale of deep neural networks has become extremely large.Correspondingly,it is large-scale data processing and computing,but now the development of electronic integrated circuits is close to the limit of Moore's Law,so it is difficult to greatly improve the computing speed of hardware,thus limiting the development of neural networks to a higher level.Compared with traditional electrical integrated circuits,due to the extremely high raw bandwidth and fast signal transmission characteristics of optical systems,people have begun to consider integrated photonics as a computing platform.Although optical chips can perform optical communication and process data carried by optical signals,perform extremely high-speed matrix-vector multiplication,and have almost no delay dispersion and crosstalk between signals,due to the weak nonlinearity of optical signals,it is difficult to perform in optical calculations.Nonlinear operations.If the neural network needs to deal with more complex problems,it is inseparable from the construction of complex nonlinear models,which seriously limits the role of optical computing in deep learning computing,and affects the practical application of optical computing.As a tool for real-time signal processing,although the chip is extremely fast,it is difficult to store.Therefore,in actual operation,the optical chip still needs the cooperation of the electronic system for neural network calculation.In this thesis,based on the premise of convolutional neural network computing applied to image recognition,the FPGA hardware support for neural network computing with optical chip is designed,that is,the research on the electrical computing part of the optoelectronic hybrid neural network.Based on the transmission mode of optoelectronic hybrid computing,plan the overall structure of the computing architecture,including the operation unit,storage mode and interconnection modules,plan the timing design of the FPGA and the structural evaluation and selection of each module.Research and split the calculation process of convolutional neural network,including sliding convolution window for multiply-add calculation,sliding pooling window for pooling operation,etc.,and optimize the operation mode.The transmission and processing of design data,including image data,weights of each layer,calculation results,feature information and other data,use the exchange mechanism of the AXI bus to interact externally,and use the internal bus to make the convolution and pooling operation units.Value calculation and configuration of internal registers.After the design of each circuit module is finally completed,the waveform simulation is carried out.