Close

Presentation

DANCE: Differentiable Accelerator/Network Co-Exploration
TimeTuesday, December 7th3:50pm - 4:10pm PST
Location3016
Event Type
Research Manuscript
Virtual Programs
Presented In-Person
Keywords
System-on-Chip Design Methodology
Topics
EDA
DescriptionThis work presents DANCE, a differentiable approach towards the co-exploration of the hardware accelerator and network architecture design.
At the heart of DANCE is a differentiable evaluator network.
By modeling the hardware evaluation software with a neural network, the relation between the accelerator architecture and the hardware metrics becomes differentiable, allowing the search to be performed with backpropagation.
Compared to the naive existing approaches, our method performs co-exploration in a significantly shorter time, while achieving superior accuracy and hardware cost metrics.