Reproduce “Chip Placement with Deep Reinforcement Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Azalia Mirhoseini, Anna Goldie, Mustafa Yazgan, Joe Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson and 14 more

A deep reinforcement learning framework, leveraging domain adaptation, effectively places chip components to optimize for power, performance, and area (PPA). This approach reduces chip placement time from weeks of human expert involvement to under 6 hours, achieving design quality that is superior or comparable to human experts and commercial tools.

Key insights

  • Deep reinforcement learning can effectively navigate the complex, high-dimensional state space of chip placement, directly optimizing non-differentiable objectives.
  • Pre-training a Graph Neural Network (GNN) on a supervised task to predict placement quality provides robust, generalizable state embeddings essential for transfer learning across diverse chip netlists.
  • Scaling the diversity and size of the pre-training dataset significantly enhances the generalization capabilities of the RL agent, leading to faster optimization and higher quality for new designs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.