A Physics-based Model of RRAM Probabilistic Switching for Generating Stable and Accurate Stochastic Bit-streams

A physics-based probabilistic switching model for RRAM stochastic number generator (SNG) is developed to obtain stable and accurate stochastic bit-streams (SBS) for stochastic computing. By taking account of the physical origin of intrinsic variations, the switching probability can be described under various operation schemes. By modeling the cumulative effect between continuous cycles, the probability shift (PS) which seriously affects the accuracy of SBS can be quantified. By using the model, an optimized Recover scheme is proposed to mitigate the PS, and obtain a wide range of stable switching probabilities. The model can be used to design the operation scheme of SNG and choose appropriate bitstream length to achieve target system performance according to the requirement of application. Furthermore, the model is successfully implemented to evaluate the performance of RRAM-based image processing system, which is a cost- and energy-efficient solution for edge detection.