Intrinsic Bounds for Computing Precision in Memristor-Based Vector-by-Matrix Multipliers

Analog computing with crossbars of memristors is a promising approach to build compact energy-efficient vector-by-matrix multiplier (VMM), a key block in many data-intensive algorithms. However, device non-linearity, process variations, interconnect parasitics, noise, and memory state drift limit the computing precision of such systems. In this article, we investigate the impact of such non-idealities in analog current-mode memristive VMMs through simulations and experiments on the most prospective passive crossbars. We show that there is an optimal tuning voltage to minimize the computation error. Furthermore, error balancing and bootstrapping are introduced as two techniques for improving the precision. It is also shown that when size of N x N crossbar is scaled up, the optimum interconnect wire conductance should increase quadratically with N to preserve the computing precision when using naive error balancing approach, and that the differential scheme is imperative for temperature insensitive operation and also to reduce the IR-drop effect.