Building a Better Random Number Generator for Stochastic Computing
Florian Neugebauer, Ilia Polian, John P. Hayes · 2017
Stochastic circuits (SCs) offer tremendous area and power-consumption benefits at the expense of computational inaccuracies. They require random num-ber sources (RNSs) to implement stochastic number generators (SNGs) for all of their inputs. It is common for an SC to have a large number of primary and auxiliary inputs. Often the associated SNGs take up as much as 80% of the entire circuit area, so sharing RNSs is a very important goal in stochastic computing. Such sharing often leads to large correlation errors that have to be resolved via costly decorrelation methods. Linear feed-back shift registers (LFSRs) are typically used as RNSs. However, we show that their deterministic and linear behavior can interfere with commonly used decorrelation methods, causing systematic computation errors, and limiting the possibilities of sharing LFSRs between SNGs. We therefore propose a novel pseudo-random number generator SBoNG for stochastic circuits that combines an LFSR with a non-linear S-box function. An SBoNG does not interfere with decorrelation and can be shared effi-ciently by multiple SNGs. Consequently, SBoNGs scale very well in SCs with large numbers of inputs.