Stochastic Computing (SC) is a computational paradigm that encodes numerical values within the probability of logic ‘1’s in a stochastic bitstream. Utilizing SC enables neural network implementations to achieve improved scalability and ultra-low hardware footprint . This work proposes SuperSIM+, an extension of the benchmarking framework SuperSim designed for neural networks using superconducting Josephson devices. By expanding its functions beyond the existing single-bit and multi-bit Process-In-Memory (PIM) designs, SuperSIM+ enables comprehensive SC-based neural network (SCNN) benchmarking. Specifically, the proposed SuperSIM+ support superconducting SCNNs by incorporating stochastic memory, probabilistic multiplication, superconducting pulse merging-based accumulation, and efficient activation using adjustable flux storage loops. Several classical neural network models trained on the MNIST and CIFAR-10 datasets are employed to validate our framework in SC simulation, including computational accuracy, hardware cost, energy consumption, and overall latency .
Key words: Stochastic Computing, System Simulations, Neural Network
Acknowledgment:
This work was supported by JST FOREST Program (Grant Number JPMJFR226W, Japan) and JSPS KAKENHI Grant Number JP22H0220 and 23K28055.