The Design And Implementation Of A Computing-in-memory Simulator Based On Memristor

Neural network algorithms have a wide range of applications in many fields in data age,but they require more computing resources than traditional algorithms.It has become more and more difficult to improve the computational efficiency by improving CMOS process.Therefore,a new hardware design is urgently needed to improve the efficiency of neuromorphic computing.Memristor is a basic circuit element newly discovered in recent years,and its non-volatility can be used as a storage component.The computing structure based on memristor crossbar can reduce communication overhead between computing structure and storage components.At the same time,the parallel characteristic of memristor crossbar makes it possible to greatly accelerate the speed of neural network computations.However,memristor components cannot be integrated and applied on a large scale.And some traditional simulation platforms cannot support simulation of new types of resistance components.Therefore,a specific memristor-oriented software simulation platform is needed to explore its application in neural network computations.According to the conductance and pulse relationship of HP laboratory memristor model,device characteristics and some non-ideal characteristics can be quantified.A strategy for expanding the range of values expressed by the memristor crossbar is proposed,by which the memristor crossbar can perform matrix operations in the range of floating-point numbers.The software simulator for memristor can be designed based on these and implemented in C++ language.The simulator implements matrix-vector multiplication calculation with O(1)computational complexity,and can accelerate core operators of neural network algorithms based on the multilayer perceptron algorithm.The structure of the simulator is divided into a three-layer structure from device level to circuit level and then to function level.First,the code simulation of underlying components can be realized based on memristor and other circuit component data of HP Labs.Then the code simulation of circuit structure is realized according to the circuit simulation framework.Finally,according to the mapping and calculation strategy of neural network algorithm on memristor crossbar,a functional level simulation is realized.In the recognition test of MNIST handwritten digit set,the simulator realizes mapping and computation of the multilayer perceptron algorithm using the memristor computing structure.Experiments show that without considering non-ideal characteristics,learning accuracy rate is as high as 92%;when considering non-ideal characteristics,learning accuracy rate will drop by 20%.Compared with the computing structure of SRAM array,time delay is reduced by an order of magnitude,and energy consumption is reduced by 80%.In the test of floating-point matrix calculation,the simulator has improved the calculation accuracy by 5-10% compared with NeuroSim.