Presentation
A Highly Efficient Reinforcement Learning Based DFG Mapping Method on CGRA
DescriptionIn this paper, we propose a reinforcement learning (RL) based DFG mapping, which learns an optimization algorithm with synthetic DFGs as a training dataset for a given CGRA architecture. As opposed to supervised learning, our approach is based on reinforcement learning which overcomes the limitation of the requirement of the large labeled dataset. Our experimental results reveal that on average our RL-based mapping approach finds an initial solution 6.45x faster and 7.01x more solution than state-of-the-art approaches. We also show that our approach can find more low-power solutions.
Event Type
Work-in-Progress Poster
TimeWednesday, July 12th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security