End-to-End Stochastic Computing

Embedded systems that require high performance hardware for signal processing, optimization, and control have historically relied on optimized fixed-point algorithms deployed on low-power microprocessors and/or digital signal processors. However, the sensing and actuation interfaces in such systems increasingly rely on oversampled, single-bit sigma-deltamodulated (SDM) representations of data inputs and control outputs. Conventional computing substrates requires conversion both on the input side and on the output side in order to effectively interface with physical systems that utilize SDM representations. This position paper argues that embedded systems should be designed to sense, process, compute, and actuate using an approach called stochastic computing, which operates directly on the oversampled SDM representation that is a natural fit for interfacing with physical systems. Stochastic computing offers numerous advantages in terms of datapath simplicity, robustness, error tolerance, and support for variable precision, all of which make it an appealing approach for these embedded applications. On the other hand, stochastic computing has a number of limitations that make design of entire systems based on this approach quite challenging. This paper outlines some of those challenges and details our experiences with building such a prototype system.