Large-scale neuromorphic computing hardware for analog AI enabled by epitaxial random access memory

Abstract

A neuromorphic computing on memristor-based crossbars is one of promising next generation analog computing methods since it features fast switching speed, extremely small cell footprint, low energy consumption for matrix-vector multiplication, capability of both storage and computing, three-dimensionality, and many analog weight steps. Although there have been intensive studies on the development of an analog memristive device and its large-scale crossbar to implement neuromorphic hardware systems for deep neural networks, only limited approaches, such as inference task, were suggested due to spatial/temporal variations and nonlinear/step-limited weight update properties. In order to address those issues, this thesis presents epitaxial random access memory and relevant techniques at material-, device-, array-, architecture-, algorithm-level. The proposed methods have great potential to improve device performance and relax the large-scale system-level requirements for analog AI computing.