A Deep Neural Network Deployment Based on Resistive Memory Accelerator Simulation

TL;DR

This paper demonstrates training a Deep Neural Network within a ReRAM-based simulation environment using CrossSim, highlighting the impact of device noise and temperature on training accuracy and energy efficiency.

Contribution

It introduces CrossSim, an API that simulates neural network training on ReRAM devices considering non-linearities and noise, with analysis on various device configurations.

Findings

ReRAM devices can effectively train neural networks with acceptable accuracy.

Device temperature and material significantly affect training outcomes.

Simulation results guide the design of energy-efficient neural accelerators.

Abstract

The objective of this study is to illustrate the process of training a Deep Neural Network (DNN) within a Resistive RAM (ReRAM) Crossbar-based simulation environment using CrossSim, an Application Programming Interface (API) developed for this purpose. The CrossSim API is designed to simulate neural networks while taking into account factors that may affect the accuracy of solutions during training on non-linear and noisy ReRAM devices. ReRAM-based neural cores that serve as memory accelerators for digital cores on a chip can significantly reduce energy consumption by minimizing data transfers between the processor and SRAM and DRAM. CrossSim employs lookup tables obtained from experimentally derived datasets of real fabricated ReRAM devices to digitally reproduce noisy weight updates to the neural network. The CrossSim directory comprises eight device configurations that operate at different temperatures and are made of various materials. This study aims to analyse the results of training a Neural Network on the Breast Cancer Wisconsin (Diagnostic) dataset using CrossSim, plotting the innercore weight updates and average training and validation loss to investigate the outcomes of all the devices.