UNICO: Efficient Unified Hardware-Software Co-Optimization For Deep Neural Networks

Bahador Rashidi, Chao Gao, Shan Lu, Wang Zhisheng, Lu Wei, SHANGLING JUI, Di Niu

Keywords: Neural accelerator optimization, Hardware-Software co-design, Hardware optimization, HW design robustness, HW design generalizability, Successive halving, Holistic time-efficient search, Multi-objective Bayesian optimization, High-fidelity search, Tensor computation

TL;DR: UNICO is a fast and high-fidelity neural accelerator HW-SW co-search solution that can find high-quality HW configurations that are generalizable to unseen DNN applications at the time of co-search.

Abstract: Specialized hardware has become an indispensable component to deep neural network acceleration. To keep up with the rapid evolution of neural networks, recently, holistic and automated solutions for jointly optimizing both hardware architectures and software mapping have been proposed. In this paper, we propose UNICO, a Unified Co-Optimization framework for hardware-software co-design, aimed at addressing the efficiency issues of vast design space exploration and the issue of overfitting to specific input neural network workloads that are facing current approaches. UNICO employs multi-objective Bayesian optimization to sample hardware, and performs parallel and adaptive software mapping search for hardware samples with a customized successive halving algorithm. To reduce overfitting, UNICO incorporates quantitative robustness measures to guide the proposed search and evaluation procedure. Experiments performed for both open-source spatial accelerators and a real-word commercial environment show that UNICO significantly improves over its counterparts, finding not only superior but also more robust hardware configurations, yet at drastically lower search cost.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)