You Only Search Once: A Fast Automation Framework for Single-Stage DNN/Accelerator Co-design
Weiwei Chen1,2,a, Ying Wang1,b, Shuang Yang1,2,c, Chen Liu1,d and Lei Zhang1,e
1Institute of Computer Technology, Chinese Academy of Sciences, Beijing, China
2University of Chinese Academy of Sciences, Beijing, China
achenweiwei@ict.ac.cn
bwangying2009@ict.ac.cn
cyangshuang2019@ict.ac.cn
dliucheng@ict.ac.cn
ezlei@ict.ac.cn
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 fails 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 userspecified energy and latency optimization constraints, respectively.
Keywords: Automl, Hardware/Software co-design, Acceleration.
