[18p-Z27-1~14] 11.5 Junction and circuit fabrication process, digital applications

△ [18p-Z27-6] Design and evaluation of an AQFP/RSFQ sigmoid-function generator for neural networks based on stochastic computing

〇(M2)Wenhui Luo¹, Naoki Takeuchi², Olivia Chen², Nobuyuki Yoshikawa^1,2 (1.Yokohama Natl. Univ., 2.IAS, Yokohama Natl. Univ.)

Keywords:superconductor, stochastic computing, sigmoid-function

We propose a sigmoid-function generator (SFG) using adiabatic quantum-flux-parametron (AQFP) circuit and rapid single-flux-quantum (RSFQ) circuit for hardware implementation of deep neural networks based on stochastic computing. The SFG corresponds to an AQFP/RSFQ interface and a superconducting loop performed as a counter. The flux quantum can be stored in loop inductance and the polarity of loop current depends on the logic in random bit stream from the former AQFP circuit. Finally, an AQFP buffer chain is used to read out the polarity to perform sigmoid-function. We designed and fabricated a sigmoid-function generator using AIST high-speed standard process (HSTP) and demonstrated the generation at 100kHz.