Extremely Energy-Efficient Non-Linear Function Approximation Framework Using Stochastic Superconductor Devices

Abstract

Recently developed adiabatic quantum-flux-parametron (AQFP) superconducting technology achieves the highest energy efficiency among various superconducting logic families, potentially 10 ⁴ -10 ⁵ gain compared with state-of-the-art CMOS. Besides ultra-high energy efficiency, AQFP exhibits two unique characteristics: the deep pipelining nature as all logic gates are clocked and the potential of building stochastic number generators (SNGs) using a single AQFP gate, far more efficient than SNGs implemented in conventional CMOS. These unique characteristics indicate that the AQFP technology is highly compatible with stochastic computing (SC) implementations, where operands are represented by a time-independent bit sequence utilizing the deep pipelining structure of AQFP. To shed some light on the SC-based design methodology on novel superconducting technologies, we propose an AQFP-based non-linear function approximation framework with the fashion of Bernstein polynomials, achieving a general hardware architecture to perform multiple non-linear functions without any extra hardware overhead. Experimental results of 9 common non-linear functions widely used in pattern recognition, signal processing, and neural networks reveal that our work provides outstanding energy efficiency with sufficient computing accuracy. The energy-delay-error-product (EDE _MAE P) of the proposed design, in terms of the polynomial degree of 3, 5 and 7, are 3.47 × 10 ⁻²⁵ J $\cdot$ s, 3.63 × 10 ⁻²⁵ J $\cdot$ s and 6.79 × 10 ⁻²⁵ J $\cdot$ s on average, respectively, achieving 5-6 orders better performance than its CMOS counterpart. Further discussions on the measurement results of trial-fabricated AQFP comparators reveal the future research directions of AQFP-based SC implementations.