SMART: A Heterogeneous Scratchpad Memory Architecture for Superconductor SFQ-based Systolic CNN Accelerators

TL;DR

This paper introduces SMART, a heterogeneous scratchpad memory architecture combining shift-register and random access memory to significantly boost throughput and reduce energy consumption in superconductor SFQ-based CNN accelerators.

Contribution

The paper presents a novel heterogeneous SPM architecture, SMART, with a new low-power, dense CMOS-SFQ random access array and an ILP-based compiler for improved CNN inference performance.

Findings

SMART achieves 3.9x throughput increase over previous SFQ accelerators.

Inference energy is reduced by 86% with SMART.

Experimental results confirm significant performance and energy efficiency improvements.

Abstract

Ultra-fast \& low-power superconductor single-flux-quantum (SFQ)-based CNN systolic accelerators are built to enhance the CNN inference throughput. However, shift-register (SHIFT)-based scratchpad memory (SPM) arrays prevent a SFQ CNN accelerator from exceeding 40\% of its peak throughput, due to the lack of random access capability. This paper first documents our study of a variety of cryogenic memory technologies, including Vortex Transition Memory (VTM), Josephson-CMOS SRAM, MRAM, and Superconducting Nanowire Memory, during which we found that none of the aforementioned technologies made a SFQ CNN accelerator achieve high throughput, small area, and low power simultaneously. Second, we present a heterogeneous SPM architecture, SMART, composed of SHIFT arrays and a random access array to improve the inference throughput of a SFQ CNN systolic accelerator. Third, we propose a fast, low-power and dense pipelined random access CMOS-SFQ array by building SFQ passive-transmission-line-based H-Trees that connect CMOS sub-banks. Finally, we create an ILP-based compiler to deploy CNN models on SMART. Experimental results show that, with the same chip area overhead, compared to the latest SHIFT-based SFQ CNN accelerator, SMART improves the inference throughput by $3.9\times$ ($2.2\times$), and reduces the inference energy by $86\%$ ($71\%$) when inferring a single image (a batch of images).