Magnetic tunnel junction-based crossbars: improving neural network performance by reducing the impact of non-idealities

Published

July 13, 2023

Author(s)

William Borders, Nitin Prasad, Brian Hoskins, Advait Madhavan, Matthew Daniels, Vasileia Gerogiou, Tiffany Santos, Patrick Braganca, Mark Stiles, Jabez J. McClelland

Abstract

Increasingly higher demand in chip area and power consumption for more sophisticated artificial neural networks has catalyzed efforts to develop architectures, circuits, and devices that perform like the human brain. However, many novel device technologies suffer from non-idealities such as device variation, or circuit sneak paths that reduce network accuracy. Here, we report that an array of magnetic tunnel junction devices integrated with complementary metal oxide semiconductors (CMOS) greatly reduces the impact of non-idealities in the circuit and performs inference with accuracies nearly identical to software.

Proceedings Title

Magnetic tunnel junction-based crossbars: improving neural network performance by reducing the impact of non-idealities

Conference Dates

May 15-19, 2023

Conference Location

Sendai, JP

Pub Type

Conferences

Download Paper

Local Download

Citation

Borders, W. , Prasad, N. , Hoskins, B. , Madhavan, A. , Daniels, M. , Gerogiou, V. , Santos, T. , Braganca, P. , Stiles, M. and McClelland, J. (2023), Magnetic tunnel junction-based crossbars: improving neural network performance by reducing the impact of non-idealities, Magnetic tunnel junction-based crossbars: improving neural network performance by reducing the impact of non-idealities, Sendai, JP, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=936143 (Accessed April 28, 2024)