A stochastic-computing based deep learning framework using adiabatic quantum-flux-parametron superconducting technology

Abstract

The Adiabatic Quantum-Flux-Parametron (AQFP) superconducting technology has been recently developed, which achieves the highest energy efficiency among superconducting logic families, potentially 10 4 -10 5 gain compared with state-of-theart CMOS. In 2016, the successful fabrication and testing of AQFP-based circuits with the scale of 83,000 JJs have demonstrated the scalability and potential of implementing largescale systems using AQFP. As a result, it will be promising for AQFP in high-performance computing and deep space applications, with Deep Neural Network (DNN) inference acceleration as an important example.Besides ultra-high energy efficiency, AQFP exhibits two unique characteristics: the deep pipelining nature since each AQFP logic gate is connected with an AC clock signal, which in AQFP, (ii) a high-accuracy sub-sampling (pooling) block in AQFP, and (iii) majority synthesis for further performance improvement and automatic buffer/splitter insertion for requirement of AQFP circuits. Experimental results suggest that the proposed SC-based DNN using AQFP can achieve up to 6.8 × 10 4 times higher energy efficiency compared to CMOS-based implementation while maintaining 96% accuracy on the MNIST dataset.

Temporal and SFQ pulse-streams encoding for area-efficient superconducting accelerators

“…The minimum distance between pulses increases with the number of inputs. counting network (M inputs, one output) where M is a power of two [4,6,31]. A balancer (Figure 6a) is the building block for counting networks.…”

“…To address SFQ constraints, Cai, et al [6], proposed a stochastic computing-based deep learning framework using adiabatic Quantum-Flux-Parametron technology. Also, Tzimpragos et al [51] adapted RL to RSFQ.…”

“…Previous work has explored unary approaches such as SSC and RL in SFQ technology, demonstrating beneficial trade-offs between area and energy/power consumption for selected applications [6,51]. Stream-based methods, such as SSC, are particularly efficient for mathematical operations yet sacrifice accuracy to latency.…”

A Computational Temporal Logic for Superconducting Accelerators

“…However, this design choice compromises the advantage speed of SFQ technology [4] as the number of execution cycles per instruction increases with the number of bit slices. Moreover, the use of shift register-based memories -given the lack of dense, fast, and robust cryogenic memory blocks -seems to be the only reasonable choice at the moment; still not a viable solution though for large-scale designs 3 .…”

“…Another promising effort is the stochastic computingbased deep learning acceleration framework presented by R. Cai, et al [3]. The authors of this work took advantage of stochastic computing's time-independent bit sequence value representation and the small hardware footprint of its operators to redesign the basic neural network components in AQFP and were able to achieve orders of magnitude energy improvements compared to CMOS.…”