How to use stochastic devices in probabilistic calculations

Many statistically-motivated scientific computing applications require sampling probability distributions. Switching from software-defined random number generators to specialized stochastic devices may not only make the computationally expensive process of sampling cheaper, but can motivate the formulation of more complex approaches that shift additional burden away from traditional computation and towards sampling. This talk focuses on establishing figures of merit for stochastic devices which are derived from the quality of the samples they produce. We will evaluate experimentally-acquired bitstreams from magnetic tunnel junctions and tunnel diodes. On the surface, the requirements for stochastic devices are daunting, but their quality can be improved using error correction and feedback-control. Their efficiency is overwhelmingly a function of how they are integrated into a circuit, and not from specific choice of the device. Finally, we go on to use the bitstream-derived samples in model calculations to show the efficacy of a hardware-based approach.