Design Of Digital Control Platform Based On Optical Neural Network

Among the deep learning algorithms,the convolutional neural network is the most representative model,which is widely used in image classification,non-linear prediction and other scenes,showing a high degree of intelligence.Operational operations in convolutional neural networks are dominated by convolution operations,which involve large-scale operations such as dense matrix multiplication,which requires a large amount of storage and computing resources during the operation.Especially when the matrix scale is large,the traditional microelectronic processor is difficult to provide sufficient computing power for the convolutional neural network,and the operation efficiency is low.In recent years,optical computing has been gradually used in the acceleration technology of convolutional neural networks due to its fast speed,parallelism and low power consumption.The core of optical calculation is the Mach-Zehnder interferometer(MZI),which forms a network structure array through MZI cascade.According to the principle of matrix singular value decomposition,when coherent light is input,the coherent transmission of light in the MZI network array can be equivalent to the realization speed Extremely fast high-order analog matrix operations.However,the development of optical computing is still not mature enough to realize the neural network of all-optical computing.It needs to be combined with a microelectronic processor as part of data flow control and non-linear computing.In a hybrid neural network composed of optical computing chips + microelectronic processors,matrix operations are performed at the speed of light.Therefore,the calculation speed bottleneck of optical neural networks is mainly the speed of the electro-optical interface,which not only requires the hardware support of high-speed DAC and ADC,but A digital signal stream with sufficient bandwidth is required for matching.This research uses FPGA development board as the data flow control part of the microelectronic processor of the photoelectric hybrid neural network.FPGA is a field programmable integrated circuit with rich hardware resources.With its powerful parallel signal processing capabilities,it is outstanding in dealing with intensive data streaming tasks.It is often used as a hardware acceleration platform for neural networks,and is similar to GPUs.It also has the advantage of low power consumption.However,FPGAs are far less flexible and efficient than ARM processors in the scheduling of complex tasks.Therefore,the use of a heterogeneous So C development platform that integrates ARM processors and FPGAs for software and hardware collaborative development can give full play to the unique advantages of the two architectures,and combines the high parallelism of FPGAs with the high flexibility of ARM processors.In addition,FPGA has a variety of powerful IP soft cores,and customized design of IP soft cores can simplify various development requirements,improve development efficiency,and accelerate the process of design closure.Finally,a digital stream heterogeneous control platform of photoelectric hybrid neural network was built on the Xilinx ZCU111 development platform,in which ARM Cortex-A53 was used as the main processor,Ultra Scale+ FPGA was used as the coprocessor,and a high-speed digital-to-analog conversion was designed based on the IP soft core.Interface,where the DAC sampling frequency is 0.98 GSPS,and the ADC sampling frequency is 1.96 GSPS.The high-bandwidth data flow path implemented in this design provides an effective digital control platform for the application of optical computing chips on neural networks.