Dot-Product Engine as Computing Memory to Accelerate Machine Learning Algorithms

Miao Hu, John Paul Strachan, Zhiyong Li, R. Stanley Williams
Hewlett Packard Labs


Abstract

Currently, intense work is underway to develop memristor crossbar arrays for high density, nonvolatile memory applications. However, another capability of memristor crossbars – natural dot-product operation for vectors and matrices – holds even greater potential for next-generation computing, including accelerators, neuromorphic computing, and heterogeneous computing. In this paper, we present a dot-product engine (DPE) based on memristor crossbars optimized for dense matrix computation, which is dominated in most machine learning algorithms. We explored multiple methods to enhance DPE’s dot-product computing accuracy. Moreover, instead of training crossbars, we try to directly use existing software-trained weight matrices on DPEs so no heroic effort is needed to innovate learning algorithms for new hardware. Our results show that computations utilizing DPEs can achieve 1000 ~ 10000 times better speed-efficiency product comparing to a state-of-art ASIC. And machine learning algorithm utilizing DPEs can easily achieve software-level accuracy on testing. Both experimental demonstrations and data-calibrated circuit simulations are presented to demonstrate the realistic implementation of a memristor crossbar DPE.