Implementation of a Binary Neural Network on a Passive Array of Magnetic Tunnel Junctions
Authors:
Jonathan M. Goodwill,
Nitin Prasad,
Brian D. Hoskins,
Matthew W. Daniels,
Advait Madhavan,
Lei Wan,
Tiffany S. Santos,
Michael Tran,
Jordan A. Katine,
Patrick M. Braganca,
Mark D. Stiles,
Jabez J. McClelland
Abstract:
The increasing scale of neural networks and their growing application space have produced demand for more energy- and memory-efficient artificial-intelligence-specific hardware. Avenues to mitigate the main issue, the von Neumann bottleneck, include in-memory and near-memory architectures, as well as algorithmic approaches. Here we leverage the low-power and the inherently binary operation of magn…
▽ More
The increasing scale of neural networks and their growing application space have produced demand for more energy- and memory-efficient artificial-intelligence-specific hardware. Avenues to mitigate the main issue, the von Neumann bottleneck, include in-memory and near-memory architectures, as well as algorithmic approaches. Here we leverage the low-power and the inherently binary operation of magnetic tunnel junctions (MTJs) to demonstrate neural network hardware inference based on passive arrays of MTJs. In general, transferring a trained network model to hardware for inference is confronted by degradation in performance due to device-to-device variations, write errors, parasitic resistance, and nonidealities in the substrate. To quantify the effect of these hardware realities, we benchmark 300 unique weight matrix solutions of a 2-layer perceptron to classify the Wine dataset for both classification accuracy and write fidelity. Despite device imperfections, we achieve software-equivalent accuracy of up to 95.3 % with proper tuning of network parameters in 15 x 15 MTJ arrays having a range of device sizes. The success of this tuning process shows that new metrics are needed to characterize the performance and quality of networks reproduced in mixed signal hardware.
△ Less
Submitted 6 May, 2022; v1 submitted 16 December, 2021;
originally announced December 2021.
Temperature overshoot as the cause of physical changes in resistive switching devices during electro-formation
Authors:
Jingjia Meng,
Bingyuan Zhao,
Jonathan M. Goodwill,
James A. Bain,
Marek Skowronski
Abstract:
Resistive switching devices based on transition metal oxides require formation of a conductive filament in order for the device to be able to switch. Such filaments have been proposed to form by the reduction of the oxide due to application of the electric field, but this report seeks to rebut that interpretation. Frequently reported physical changes during electro-formation include delamination o…
▽ More
Resistive switching devices based on transition metal oxides require formation of a conductive filament in order for the device to be able to switch. Such filaments have been proposed to form by the reduction of the oxide due to application of the electric field, but this report seeks to rebut that interpretation. Frequently reported physical changes during electro-formation include delamination of electrodes, crystallization of functional oxide, intermixing of electrode and oxide materials, and extensive loss of oxygen presumably to the ambient. Here we show that most of these effects are not inherent to the formation and switching processes and instead are due to an experimental artifact: the discharge of parasitic capacitances in the forming circuit. Discharge of typical BNC cables can raise the temperature of the filament to between 2,000 and 5,000 K resulting in extensive physical changes. Discharge and associated effects mentioned above can be eliminated using an on-chip load element without affecting the ability to switch.
△ Less
Submitted 14 April, 2020;
originally announced April 2020.
