An Energy Efficient Multi-Rail Architecture for Stochastic Computing: A Bayesian Sensor Fusion Case Study

International audienceRecently, Stochastic Computing has sparked interest in Bayesian inference resolution for its promising efficiency in area and power consumption. This representation encodes values by the rate of bits at '1’ in a bit-stream. Still, in a sequential  architecture, most of the energy cost is due to the long computation time required for achieving a  satisfying accuracy. In this paper, we propose a multi-rail architecture for Bayesian sensor fusion problems based on a Shift Register Isolator and permutations in order to reduce the computation time and thus, the energy consumption, without a significant increase in area. Indeed, with this resource sharing strategy, we are able to reduce the energy consumption by up to 73% in return for an area overhead of 24%, while  maintaining the computation accuracy.