Characterizing Stochastic Number Generators for Accurate Stochastic Computing

Since the advent of the post-Moore era, the challenges of increasing power have been raised. Stochastic computing, where stochastic numbers or binary bit streams are used to convey information, adopts encoding method that is different from the conventional multi-bit binary radix representation. This encoding method enables low arithmetic hardware cost, reduced energy consumption, and improved reliability. Despite the aforementioned advantages, the accuracy of a stochastic computing system is largely affected by the stochastic number generators (SNGs), which converts binary numbers into their stochastic representation. This paper reviews recent developments in SNGs and their implementations. The accuracy of the computation results produced by these SNGs is evaluated and compared.