An Industrial Perspective on ML-Macro Placement Methods

The advent of Machine Learning (ML) and Reinforcement Learning (RL) augmented macro placement techniques have shown to improve key design metrics on synthetic designs, within a few hours. This ML-driven approach to a traditionally iterative process typically spanning multiple weeks has architected an industrial mindset change towards automated macro placement. However, can the techniques replicate the impressive results on real-world designs, thereby bridging the gap between research and industry? This paper explores the answer to the above question by experimenting with ML-driven macro placement techniques from a leading external EDA tool on industry leading GPU designs. We define the bounding scope, document the process, publish the results, and analyze the experiments to provide an industrial perspective to the existing ML-macro placement methods. This industrial perspective forms the crux of this paper detailing the different implementation challenges and providing recommendations so that existing ML-methods can be fine-tuned to bridge the gap between research and industry needs.