Research On Processing-in-memory Architecture For Neural Network Computation In ReRAM-based Main Memory

Metal-oxide resistive random-access memory(Re RAM)and its associated crossbar array structure have provided the basis for promising architecture for NN computing circuits and system.However,due to the scarcity of neural network algorithm mapp ing methods,slow circuit-level simulation speeds,and immature device processes that affect data accuracy,Re RAM-based neural network computing systems cannot efficiently implement large-scale neural network algorithms.Aiming at the above problems,this article uses top-down research ideas to carry out research on the neural network architecture of integrated storage and computing.In the design from the algorithm level to the structure level,this paper proposes a method of multi-dimensional convolution layer operation using Re RAM array based on the mapping method of matrix-vector multiplication.Aiming at the problem of setting the weight value accuracy,this dissertation proposed a multi-bit weights method based on positive and negative weights to reduce the high energy and area consumptions.Based on the algorithm mapping structure,this dissertation proposes a static operating point analysis method to simplify the nonlinear equations of Re RAM array calculation.The array scale,device process,and neural network architecture are used as variable parameters.A parametric structure description method of Re RAM neural network computing system and a behavior-level simulation model that simulates Re RAM array calculation accuracy are proposed to improve simulation speed.In the design from the circuit level to the device level,this dissertation proposes the use of a three-dimensional MNIST edge detection method to verify the reliability of a three-dimensional resistive memory array based on a field programmable gate array(FPGA)-controlled relay matrix test platform.The convolutional neural network computing system architecture lays the foundation.The research in this dissertation demonstrates that it is necessary to carry out collaborative optimization of algorithms and hardware in order to make full use of the advantages of Re RAM neural network computing systems in intelligent applications.