AGD: A Learning-based Optimization Framework for EDA and its Application to Gate Sizing
DescriptionIn EDA, most simulation models are not differentiable, and many design decisions are discrete. As a result, greedy optimization based on numerical gradients have been used widely, although it suffers from suboptimal solutions. Reinforcement learning has been leveraged to tackle this problem; however, applying RL to EDA also suffers from notorious sample inefficiency. This paper proposes an alternative to RL for EDA. We demonstrate that our method can be very close to an industry-leading commercial tool in terms of QoR of gate sizing, while it only takes several person-months in comparison to dedicated efforts over decades to develop.
Event Type
Research Manuscript
TimeThursday, July 13th2:10pm - 2:25pm PDT
Location3002, 3rd Floor
RTL/Logic Level and High-level Synthesis