Simulating Pattern Recognition Using Non-volatile Synapses: MRAM, Ferroelectrics and Magnetic Skyrmions

TL;DR

This paper investigates the use of non-volatile spintronic synapses, including MRAM, ferroelectrics, and skyrmions, in neuromorphic computing for pattern recognition, through comprehensive simulation of various models.

Contribution

It introduces a simulation framework for evaluating different non-volatile spintronic synapses in neural networks, highlighting their potential for energy-efficient AI hardware.

Findings

Spintronic synapses enable low-power neuromorphic computing.

Simulation results compare MRAM, ferroelectrics, and skyrmions for pattern recognition.

Non-volatile synapses show promising performance for future AI hardware.

Abstract

This project explores the use of non-volatile synapses in neuromorphic computing for pattern recognition tasks through a comprehensive simulation-based approach. The main approach is through spintronic synapses, which leverage the electron's spin properties to achieve efficient data processing and storage. This offers a promising alternative to traditional electronic synapses which require constant power recharge to prevent data leakage. The goal is to develop and simulate a neural network model that incorporates spintronic synapses, examining their potential to perform complex pattern recognition tasks such as image and sound classification. By building a simulation environment, we will replicate various models, including spin transfer torque based MRAM, voltage controlled magnetic anisotropy based MRAM, ferroelectric field effect transistors, and skyrmion based nanotrack for synaptic devices, to evaluate their performance and compare results across different non-volatile implementations. The findings will highlight the effectiveness of spintronic synapses in creating low-power, high-performance neuromorphic hardware, providing valuable insights into their application for future energy-efficient artificial intelligence systems.