FlowPlace: AI-Driven Chip Placement Using Flow Matching
FlowPlace has unveiled a flow matching model designed for chip placement, overcoming the shortcomings of diffusion-based techniques. This model incorporates mask-guided synthetic data creation, efficient training through flow-based methods with adaptable prior injection, and sampling of hard constraints to ensure layouts are free of overlaps. Evaluated against the OpenROAD and ICCAD 2015 benchmarks, FlowPlace demonstrates superior power, performance, and area (PPA) metrics, achieving sampling efficiency that is 10-50× faster and eliminating overlaps entirely. This research is available on arXiv in the computer science hardware architecture category.
Key facts
- FlowPlace uses flow matching for chip placement.
- It overcomes limitations of diffusion models: random synthetic pre-training, long sampling times, and overlap issues.
- Features include mask-guided synthetic data generation and hard constraint sampling.
- Tested on OpenROAD and ICCAD 2015 benchmarks.
- Achieves better PPA metrics, 10-50× faster sampling, and zero overlaps.
- Published on arXiv (2604.23658) in computer science hardware architecture.
Entities
Institutions
- arXiv