Triple-Rail Stochastic Number and Its Applications

Abstract

In Stochastic Computing (SC), we use bit-strings called stochastic numbers (SNs) to perform calculations in very low-cost hardware modules. In SC the length of SNs should be very long, so we need enough precision; the necessary cycles to perform calculations with SNs should be long. To mitigate this disadvantage, there proposed a method to use two shorter bit-strings to represent a single SN; the two bit-strings are called a dual-rail SN, and it is reported that it can reduce the necessary cycles with some additional hardware overhead and some calculation errors. In this paper, we further seek this direction, i.e., we propose to use three bit-strings called triple-rail SNs which needs much shorter cycles than dual-rail and conventional SNs. We compare triple-rail, dual-rail and conventional SNs in terms of hardware overhead and calculation errors.