Evaluation of memristor models for large crossbar structures

Abstract

This paper is focused on comparing selected SPICE models of TiO2 memristors with respect to time- and memory requirements in the simulation of very large artificial neural networks, which are most likely the first real-world applications of memristors as analog memories. All models were implemented as HSPICE macros and simulated in a Multilayer Perceptron artificial neural network with variable configuration. The results show that after applying modifications to the models in order to prevent numerical overflows it is possible to simulate networks with tens of thousands of memristors.