Reducing the Overhead of Stochastic Number Generators Without Increasing Error

Abstract

Stochastic Computing (SC) is an approximation method to calculate functions with very low hardware cost. When we calculate f(x) by SC, we need to generate many Stochastic Numbers (SNs) whose values are equal to x and constant values. To generate one SN, we need a linear-feedback shift register (LFSR) and a comparator, and thus such overhead may diminish the advantage of SC. Thus, there has been proposed an efficient method to duplicate x (for calculating a power of x) with low hardware overhead. However, it is not known how to reduce the hardware overhead for generating many constant SNs. Therefore, in this paper, we propose an efficient scheme to generate constant SNs with reasonably low hardware overhead without increasing errors. We provide some experimental results by which we can confirm that our proposed scheme is in general very useful to reduce the hardware overhead for generating constant SNs without increasing errors.