Methods for reducing memristor crossbar simulation time

Abstract

Memristor crossbars have the potential to perform parallel resistive computations in the analog domain, and they can be used to develop high density neural network algorithms. However, accurately simulating large memristor crossbars in SPICE (with more than 256 devices) is very difficult and time consuming. This paper discusses using Xyce (a parallel SPICE platform developed by Sandia Labs) to speed up memristor crossbar simulation. Using Xyce, we were able to successfully train neuromorphic memristor crossbars containing 10,096 memristors to learn a large array of linearly separable logic functions. Large memristor crossbars were also used for pattern recognition using both the MNIST and CBCL face datasets. To further reduce training time, a memristor crossbar approximation was simulated in MATLAB. Modeling a crossbar in MATLAB takes significantly less time, but is slightly less accurate. The trained resistance values determined by MATLAB were then downloaded to the more precise crossbar simulated in Xyce (which contains input drivers, comparator circuits, and wire resistance). The classification accuracy found in Xyce was then compared to the accuracy determined when testing the approximated crossbar in MATLAB, as well as a traditional software neural network implementation. To the best of our knowledge, this is the first published result that describes using XYCE to simulate a neuromorphic memristor crossbar using accurate memristor modeling techniques.