In-memory neural network accelerator based on phase change memory (PCM) with one-selector/one-resistor (1S1R) structure operated in the subthreshold regime

Abstract

In-memory computing (IMC) shows a disruptive potential for accelerating artificial intelligence (AI) in both inference and training tasks. Scalable IMC, however, requires novel memory technologies with extremely low current. Here we demonstrate ultra-low current matrix-vector multiplication (MVM) in a crosspoint array of phase change memory (PCM) and ovonic threshold switch (OTS) with one-selector/one-resistor (181R) structure operated in the subthreshold regime. Thanks to highly-uniform sub-$\mu$A currents, the 181R PCM crosspoint array rejects parasitic IR drop across wires, enabling excellent scaling compared to other memory devices. Our simulation of a fullyconnected neural network (FCNN) with ternary weights indicates an accuracy of 98% for MNIST classification with an array size of 512x512, which strongly supports subthreshold-operated 181R crosspoint arrays for neural network inference accelerators.