You Only Search Once: A Fast Automation Framework for Single-Stage DNN/Accelerator Co-design

TL;DR

YOSO is a single-stage automation framework that efficiently co-designs DNNs and accelerators, optimizing for accuracy, energy, and latency simultaneously, outperforming traditional two-stage methods.

Contribution

The paper introduces YOSO, a novel single-stage framework that jointly optimizes DNN and hardware design for better QoR and performance.

Findings

Achieves 1.42x~2.29x energy reduction at same precision

Achieves 1.79x~3.07x latency reduction

Outperforms two-stage co-design methods on baseline systolic array and Cifar10

Abstract

DNN/Accelerator co-design has shown great potential in improving QoR and performance. Typical approaches separate the design flow into two-stage: (1) designing an application-specific DNN model with high accuracy; (2) building an accelerator considering the DNN specific characteristics. However, it may fail in promising the highest composite score which combines the goals of accuracy and other hardware-related constraints (e.g., latency, energy efficiency) when building a specific neural-network-based system. In this work, we present a single-stage automated framework, YOSO, aiming to generate the optimal solution of software-and-hardware that flexibly balances between the goal of accuracy, power, and QoS. Compared with the two-stage method on the baseline systolic array accelerator and Cifar10 dataset, we achieve 1.42x~2.29x energy or 1.79x~3.07x latency reduction at the same level of precision, for different user-specified energy and latency optimization constraints, respectively.