Multilevel memristor based matrix-vector multiplication: influence of the discretization method

A study on the use of memristors with discrete multilevel capability for the implementation of matrix-vector multiplication (MVM) is presented. We focus our analysis on the role of the discretization methodology employed to translate the matrix coefficients into the memristor device conductive levels. The implications on the MVM configurations are analysed in the context of hardware neural networks, oriented to the optimization of neuromorphic computing algorithms.