Optimization and Modeling of Transposable Memristor Crossbars for Low Power On-Chip Learning
Abstract
With the massive impact of AI, energy consumption of required computing resources is trending toward unmanageable. Moving the computation load of these systems from data centers to more novel compute-in-memory hardware would provide vast efficiency gains. Thus, we propose a memristor crossbar circuit for neural network layers, that is reliable, transposable, easily programmable, and extremely energy efficient. We demonstrate the effectiveness of neural networks based on these crossbars first using SPICE at the circuit component level. Then we advance toward larger neural networks, using a simulation approach that can capture circuit level detail (including parasitic resistance, wire resistance, and leakage current) that is able to simulate at speeds orders of magnitude faster than a pure SPICE approach. We then use our simulation tools to discuss the energy requirements of these memristor based neural networks, including tradeoffs between energy, power, area, time, and accuracy. While we demonstrate the effectiveness of this memristor crossbar-based learning system using backpropagation through fully connected layers, this technique can be applied to any algorithm requiring bi-directional vector-matrix multiplication.