Low-Energy Spiking Neural Network Using Ge4Sb6Te7 Phase Change Memory Synapses

Abstract

Spiking neural networks (SNNs) with phase change memory (PCM) devices are promising for in-memory computing, but their performance is often constrained by the abrupt depression (conductance decrease) of PCM synapses. Here, we report an energy-efficient SNN using Ge ₄ Sb ₆ Te ₇ (GST467) as the phase-change material in a single device per synapse with identical pulses, and find $\sim 2.5\times $ reduction of inference energy in such SNNs compared to SNNs using two conventional PCMs per synapse. We leverage the unique gradual potentiation and depression characteristics of Ge ₄ Sb ₆ Te ₇ in a behavioral model and train a two-layer SNN to demonstrate both pattern and online learning. We also uncover the trade-offs between energy consumption and SNN recognition rate considering resistance drift and conductance ranges of the synapses, providing a design guideline for future energy-efficient PCM-based SNN.