Analog Weights in ReRAM DNN Accelerators

Abstract

Artificial neural networks have become ubiquitous in modern life, which has triggered the emergence of a new class of application specific integrated circuits for their acceleration. ReRAM-based accelerators have gained significant traction due to their ability to leverage in-memory computations. In a crossbar structure, they can perform multiply-and-accumulate operations more efficiently than standard CMOS logic. By virtue of being resistive switches, ReRAM switches can only reliably store one of two states. This is a severe limitation on the range of values in a computational kernel. This paper presents a novel scheme in alleviating the single-bit-per-device restriction by exploiting frequency dependence of v-i plane hysteresis, and assigning kernel information not only to the device conductance but also partially distributing it to the frequency of a time-varying input. We show this approach reduces average power consumption for a single crossbar convolution by up to a factor of x16 for an unsigned 8-bit input image, where each convolutional process consumes a worst-case of 1.1mW, and reduces area by a factor of x8, without reducing accuracy to the level of binarized neural networks. This presents a massive saving in computing cost when there are many simultaneous in-situ multiply-and-accumulate processes occurring across different crossbars.