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

Published

August 6, 2024

Author(s)

Huairuo Zhang, Albert Davydov, Shafin Hamid, Eric Pop, Asir Khan

Abstract

Spiking neural networks (SNN) with Ge2Sb2Te5 (GST) based phase change memory (PCM) synaptic devices are promising for edge applications. However, from a de-vice standpoint, the performance of such SNN is often con-strained by the abrupt depression (decrease of conduct-ance) in traditional GST based PCM synapse. To overcome this, non-identical pulses or two-PCM per synapse are uti-lized, however, at the expense of an increased energy, area, and complexity for large-scale systems. Here, we report an energy-efficient SNN using Ge4Sb6Te7 (GST467) as the phase-change material in a single-PCM per synapse with identical pulses and find > 2× reduction of inference energy in such SNN compared to its two-PCM per synapse coun-terpart. We leverage the unique gradual potentiation and depression characteristics of GST467 PCM in a behavioral model and train a two-layer SNN to demonstrate both pat-tern and online learning. We also uncover the trade-offs be-tween energy consumption and SNN recognition rate con-sidering resistance drift, and conductance ranges of the synapses, providing a design guideline for future energy-efficient PCM-based SNN.

Citation

Electron Device Letters

Pub Type

Journals

Download Paper

Local Download

Keywords

phase change memory, spiking neural net-work, online learning, resistance drift

Citation

Zhang, H. , Davydov, A. , Hamid, S. , Pop, E. and Khan, A. (2024), Low Energy Spiking Neural Network using Ge4Sb6Te7 Phase-Change Memory Synapses, Electron Device Letters, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=957911 (Accessed January 17, 2025)