A Fast Converging Normalization Unit for Stochastic Computing

Abstract

Stochastic computing is a promising technology for low hardware cost and low power digital signal processing (DSP) systems. However, the quite slow convergence of stochastic normalization leads to low throughput of stochastic computing based DSP systems. In this brief, we propose a fast converging stochastic normalization unit based on joint probability tracking (JPT) method, which operates modified probability tracking on random bit streams to perform the normalization. Simulation results show that the convergence speed of the JPT scheme is as 3.5 times as that of existing methods. According to the synthesis results on 65-nm CMOS technology, the JPT scheme achieves higher (2.8 times) hardware efficiency gains with respect to the existing stochastic computing-based schemes.