Analytic variability study of inference accuracy in RRAM arrays with a binary tree winner-take-all circuit for neuromorphic applications

Abstract

With the rise of neuromorphic computing, oxide resistive RAM (OxRRAM) has received a lot of attention as potential electronic synapse. A frequently used function in neuromorphic applications is the computationally powerful winner-take-all (WTA) operation. Implementing synapses with OxRRAM devices and implementing the WTA operation with a dedicated circuit introduces an inaccuracy in obtaining the winning neuron due to OxRRAM variability and comparator offset. While winner-take-all neural networks are often used, a comprehensive variability analysis of those networks is still missing. In this work, a complete device-circuit-algorithm analytic analysis framework is developed to assess the WTA accuracy in neural networks. This framework is demonstrated on a multi-layer perceptron in OxRRAM followed by the WTA circuit. It is quantified how the classification accuracy degrades with increasing device and circuit variability. The framework sets out the boundary conditions for reliably using WTA circuits for typical OxRRAM devices. The analytic expressions in this paper are generally valid for all types of RRAM.