Research map / Image generation / Fast diffusion sampling
Chip Placement with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains a diffusion model on a novel synthetic netlist/placement dataset and uses backwards universal guidance with legality and HPWL potentials to place macros zero-shot on real IBM and ISPD circuits.
Problem
Macro placement in chip design is slow and poorly generalizing when solved with per-circuit reinforcement learning, and real placement training data is scarce and proprietary.
Approach
Replace RL with a diffusion model that places all objects simultaneously, pre-trained only on a fast synthetic netlist-and-placement generator (edges sampled with exponentially decaying probability in distance) and steered at test time by differentiable legality and HPWL guidance; the denoiser interleaves GNN and attention layers.
Key results
- On the macro-only IBM benchmark, the model reaches average HPWL 2.49e5 versus 7.33e5 for ChiPFormer and 8.72e5 for MaskPlace, with congestion 196 vs 324 for the best baseline.
- On clustered IBM, guided sampling raises legality from 0.8835 to 0.9970 and cuts HPWL by 7.1%, reaching about 20% lower HPWL than DREAMPlace.
- On mixed-size IBM placement with DREAMPlace 4.1 for standard cells, the method averages HPWL 22.7e6 versus 23.6e6 for DREAMPlace, a 4% improvement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply guided sampling to Large+v2 | Large+v2 · Clustered IBM benchmark | legality: 0.8835 -> 0.9970 | Helped |
| Apply guided sampling to Large+v2 | Large+v2 · Clustered IBM benchmark | HPWL ($10^7$): 3.203 -> 2.976 | Helped |
| Train the model on the broad-distribution v1 and v2 datasets instead of narrow v0 | Large diffusion model · Clustered IBM benchmark | legality: 0.7794 -> 0.8835 | Helped |
| Train Large using the proposed synthetic data instead of Flora's dataset | Large diffusion model · Clustered IBM benchmark | legality: 0.2640 -> 0.8213 | Helped |
| Replace the strongest baseline with Diffusion (Ours) | Macro-only placement · IBM macro-only average | HPWL ($10^5$): 7.33 -> 2.49 | Helped |
| Replace the strongest baseline with Diffusion (Ours) | Macro-only placement · ISPD macro-only average | HPWL ($10^5$): 116 -> 45.9 | Helped |
| Use Diffusion macro placements instead of DREAMPlace | Mixed-size placement · IBM mixed-size average | HPWL ($10^6$): 23.6 -> 22.7 | Helped |
| Remove attention layers (ablation) | Medium model · Synthetic v1 | legality: 0.960 -> 0.799 | Hurt |
| Remove MLP blocks (ablation) | Medium model · Synthetic v1 | HPWL Ratio: 1.039 -> 1.060 | Hurt |
| Remove sinusoidal position encodings (ablation) | Medium model · Synthetic v1 | legality: 0.960 -> 0.949 | Hurt |
| Increase model size from Small to Large | Diffusion model · Synthetic v1 | legality: 0.948 -> 0.976 | Helped |
| Use different edge-length probability functions while matching mean edge length | Diffusion model · Synthetic circuits with alternative edge distributions | Legality: 0.982; 0.983; 0.982 | No clear effect |
Surprising findings
- The model trained only on synthetic data transfers zero-shot to real IBM circuits, and the Flora-style synthetic data yields legality of only 0.264 versus 0.821 for the authors' v1 data.
Limitations
- Synthetic data does not capture multimodal edge and object-size distributions of real circuits.
- Models generalize poorly to circuits much larger than their training data, so v2 fine-tuning on larger circuits is needed.
- Mixed-size results rely on DREAMPlace for standard-cell placement, and the macro-only comparison uses re-run baselines with all macros placed.
Experimental setup
Datasets: synthetic netlist-placement data v0 synthetic netlist-placement data v1 synthetic netlist-placement data v2 ICCAD04 (IBM) benchmark ISPD2005 benchmark
Benchmarks: ICCAD04 (IBM) ISPD2005
Metrics: legality HPWL HPWL ratio RUDY congestion runtime
Baselines: MaskPlace WireMask-BBO ChiPFormer EfficientPlace DREAMPlace Flora
Builds on: DDPM universal guidance GATv2 Flora
Closest papers by meaning
- Reinforcement Learning within Tree Search for Fast Macro Placement (ICML 2024 · similarity 0.92)
- Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer (NeurIPS 2024 · similarity 0.92)
- EGPlace: An Efficient Macro Placement Method via Evolutionary Search with Greedy Repositioning Guided Mutation (ICML 2025 · similarity 0.88)
- See it to Place it: Evolving Macro Placements with Vision Language Models (NeurIPS 2026 · similarity 0.88)
- Benchmarking End-To-End Performance of AI-Based Chip Placement Algorithms (NeurIPS 2025 · similarity 0.87)
- DeepLayout: Learning Neural Representations of Circuit Placement Layout (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.