On the Reliability of Computing-in-Memory Accelerators for Deep Neural Networks

TL;DR

This paper examines the reliability challenges of computing-in-memory accelerators using non-volatile memory for deep neural networks, highlighting device issues and mitigation strategies to improve accuracy.

Contribution

It provides a comprehensive analysis of NVM device reliability issues in nvCiM DNN accelerators and reviews methods to mitigate their impact on model accuracy.

Findings

Reliability issues cause accuracy degradation in nvCiM DNNs.

Different NVM types exhibit distinct reliability challenges.

Mitigation techniques can improve the robustness of nvCiM accelerators.

Abstract

Computing-in-memory with emerging non-volatile memory (nvCiM) is shown to be a promising candidate for accelerating deep neural networks (DNNs) with high energy efficiency. However, most non-volatile memory (NVM) devices suffer from reliability issues, resulting in a difference between actual data involved in the nvCiM computation and the weight value trained in the data center. Thus, models actually deployed on nvCiM platforms achieve lower accuracy than their counterparts trained on the conventional hardware (e.g., GPUs). In this chapter, we first offer a brief introduction to the opportunities and challenges of nvCiM DNN accelerators and then show the properties of different types of NVM devices. We then introduce the general architecture of nvCiM DNN accelerators. After that, we discuss the source of unreliability and how to efficiently model their impact. Finally, we introduce representative works that mitigate the impact of device variations.