Memristor-based AI Hardware for Reliable and Reconfigurable Neuromorphic Computing

Abstract

In the field of artificial intelligence hardware, a memristor has been proposed as an artificial synapse for creating neuromorphic computer applications. Changes in weight values in the form of conductance must be identifiable and uniform to train a neural network in memristor arrays. Because of the high mobility of metal ions in the Si switching medium, an electrochemical metallization (ECM) memory has shown a high analogue switching capacity. However, switching unpredictability is caused by the extreme stochasticity of ion transport. I demonstrated a Si memristor with alloyed conduction channels that works dependably and enables large-scale crossbar array deployment. In addition, heterogeneously integrated neuromorphic chips have been developed to allow physically reconfigurable neuromorphic computing. This thesis examines alloyed metal-based silicon memristors and stackable neuromorphic chips with heterogeneous integration for reliable and reconfigurable neuromorphic computing.