Research On Convolutional Neural Network Computation Based On Memristor Crossbar Array

With the continuous development of deep learning,the structure of neural networks is becoming more and more complex and the weight parameters are increasing exponentially.The huge amount of computation puts a higher demand on the computing power.The computing hardware based on the von Neumann architecture faces the von Neumann bottleneck problem caused by the storage-compute separation structure.Memristors can fundamentally break through the von Neumann bottleneck because of their memory-computing integration.Furthermore,memristors have the potential to become the core device in artificial neural network computing due to their high computational parallelism,low power consumption,and high-density integration.Convolutional neural networks are widely used and play an important role in large-scale image recognition among various neural network models.Therefore,it is imperative to conduct research on memristor-based convolutional neural network computing.In the computation of memristor-based convolutional neural networks,the conductance state setting of memristor-cells in memristor crossbar array devices is crucial.Currently,the conductance setting of the memristor-cell depends mainly on the transistor in series with each memristor-cell in the array.The conductance of the memristor-cell is regulated by the gate voltage of the transistor.However,the introduction of the transistor transforms the memristor array device from a passive device to an active device,which leads to increased power consumption and complexity of peripheral control circuits,and reduces the integration of the devices.It is imperative to investigate an effective conductance setting strategy that eliminates the dependence of the conductance setting process on the transistor.This will enable the realization of convolutional neural network calculation based on the passive memristor crossbar array device.This thesis investigates the conductance state setting strategy of the memristor crossbar array device and designs a peripheral control circuit based on the memristor crossbar array device to implement convolutional neural network computation.The functional verification of the control circuit is completed.The main research work includes the following two aspects:(1)This study investigates the strategy for setting the conductance state of memristor-cells in an array,including both single memristor-cell and multiple memristor-cells settings.The effect of two write voltage strategies on the conductance state of a single memristor-cell is investigated,and a successive approximation conductance state setting strategy is proposed and established.Additionally,the impact of different bias schemes on the conductance state of multiple memristor-cells in the array is analyzed.Through a comparison of the 1/3 bias scheme and 1/2 bias scheme,it is determined that the 1/2 bias scheme is adopted for setting the conductance state of memristor-cells in an array.The study also investigates the influence of two conductance setting orders on the conductance state of the memristor-cells in the array,ultimately determining that the "small first and then large" order is optimal for the conductance state setting of the memristor array.Finally,the retention characteristics of the conductance state after the setting process is completed are investigated,and the setting strategy is optimized based on the research results.(2)The peripheral control circuit of memristive crossbar array device for convolutional neural network computing is investigated and designed.The study involves the design of data flow and subsequent development of a hardware architecture.Then the peripheral control circuit is built,and its function verification is realized.