Accuracy and Resiliency of Analog Compute-in-Memory Inference Engines

TL;DR

This paper evaluates the scalability of analog compute-in-memory architectures for deep neural networks, demonstrating that specialized training methods and representations can significantly improve accuracy despite device uncertainties.

Contribution

It introduces a simulation framework for large-scale DNNs with analog NVM CIM and proposes a Hessian-Aware SGD training algorithm with floating-point DNN representations for improved accuracy.

Findings

Analog NVM CIM causes high uncertainty in DNN inference.

High-precision trained DNNs are not resilient to analog NVM uncertainties.

The proposed methods improve DNN accuracy without hardware cost increase.

Abstract

Recently, analog compute-in-memory (CIM) architectures based on emerging analog non-volatile memory (NVM) technologies have been explored for deep neural networks (DNN) to improve energy efficiency. Such architectures, however, leverage charge conservation, an operation with infinite resolution, and thus are susceptible to errors. The computations in DNN realized by analog NVM thus have high uncertainty due to the device stochasticity. Several reports have demonstrated the use of analog NVM for CIM in a limited scale. It is unclear whether the uncertainties in computations will prohibit large-scale DNNs. To explore this critical issue of scalability, this paper first presents a simulation framework to evaluate the feasibility of large-scale DNNs based on CIM architecture and analog NVM. Simulation results show that DNNs trained for high-precision digital computing engines are not resilient against the uncertainty of the analog NVM devices. To avoid such catastrophic failures, this paper introduces the analog floating-point representation for the DNN, and the Hessian-Aware Stochastic Gradient Descent (HA-SGD) training algorithm to enhance the inference accuracy of trained DNNs. As a result of such enhancements, DNNs such as Wide ResNets for the CIFAR-100 image recognition problem are demonstrated to have significant performance improvements in accuracy without adding cost to the inference hardware.