Research map / Architectures and efficiency / RNNs and normalization
EGPlace: An Efficient Macro Placement Method via Evolutionary Search with Greedy Repositioning Guided Mutation
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.
Proposes EGPlace, an evolutionary macro placement method whose mutation operator scores and greedily repositions poorly placed modules, plus a linear-time overlap mask, reducing HPWL and runtime versus WireMask-EA and EfficientPlace.
Problem
Macro placement in integrated circuit design requires minimizing wirelength under overlap and congestion constraints; existing hybrid methods (e.g., WireMask-EA) are slow due to separated adjustment/reconstruction, random mutations, and costly mask computation.
Approach
An evolutionary (population-based) search whose mutation selects modules for repositioning with probability proportional to impact scores (wirelength, RUDY congestion, overlap) and reinserts them via a mask-guided greedy placement, integrating reconstruction into mutation; plus an overlap mask computed in O(iB) instead of O(iB^2).
Key results
- EGPlace reduces HPWL by 9.3% versus EfficientPlace and about 10.8% versus WireMask-EA across ISPD2005 (per the abstract/main text).
- EGPlace completes 1,000 iterations in an average of 0.81 hours, 2.8x faster than EfficientPlace and 7.8x faster than WireMask-EA.
- On ICCAD2015 (256 modules), EGPlace reduces HPWL by 49.5%, 17.4%, and 7.4% versus DreamPlace, WireMask-EA, and EfficientPlace respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EGPlace instead of EfficientPlace | Macro placement · ISPD2005 benchmarks | HPWL: 9.3% reduction in HPWL | Helped |
| Use EGPlace instead of WireMask-EA | Macro placement · ISPD2005 benchmarks | runtime: 7.8× faster | Helped |
| Use EGPlace instead of EfficientPlace | Macro placement · ISPD2005 benchmarks | runtime: 2.8× faster | Helped |
| Use EGPlace instead of EfficientPlace | Macro placement · adaptec3 | HPWL: 56.35 ± 0.99 -> 61.09 ± 1.00 | Hurt |
| Use EGPlace instead of EfficientPlace | Macro placement · bigblue1 | HPWL: 2.14 ± 0.01 -> 2.24 ± 0.03 | Hurt |
| Use EGPlace instead of EfficientPlace | Macro placement · ariane | overlap rate: 3.37 -> 1.24 | Helped |
| Use EGPlace instead of EfficientPlace | Macro placement · ariane | HPWL: 12.47 -> 7.91 | Helped |
| Add greedy repositioning rather than choosing a random valid location (ablation) | EGPlace · adaptec2 | normalized fitness | Helped |
| Add module ordering by size and connectivity rather than random module selection (ablation) | EGPlace · adaptec2 | normalized fitness | Helped |
| Add quality-guided module removal rather than uniform removal probability (ablation) | EGPlace · adaptec2 | normalized fitness | Helped |
| Use population sizes between 3 and 6 rather than too-small or too-large populations (ablation) | EGPlace · adaptec2 | layout quality | Mixed |
| Increase the module removal ratio (ablation) | EGPlace · adaptec2 | HPWL | Helped |
Surprising findings
- Crossover was dropped because experiments showed it often disrupts critical local optimization patterns in circuit placement.
Limitations
- Does not sufficiently account for mixed-size placement; mixed-size results remain inferior to DreamPlace 4.1.0 alone.
- Optimizes surrogate metrics (HPWL, RUDY) rather than final PPA objectives.
- Large benchmarks (bigblue2, bigblue4) evaluated only on a subset of 1,024 macros.
Experimental setup
Datasets: ISPD2005 Ariane RISC-V CPU ICCAD2015
Benchmarks: ISPD2005 Ariane ICCAD2015
Metrics: HPWL RUDY overlap rate runtime
Baselines: WireMask-EA EfficientPlace MaskPlace Chipformer DreamPlace GraphPlace DeepPR NTUPlace3 RePlace SP-SA
Builds on: WireMask-EA MaskPlace EfficientPlace ePlace
Closest papers by meaning
- See it to Place it: Evolving Macro Placements with Vision Language Models (NeurIPS 2026 · similarity 0.89)
- Reinforcement Learning within Tree Search for Fast Macro Placement (ICML 2024 · similarity 0.88)
- Chip Placement with Diffusion Models (ICML 2025 · similarity 0.88)
- Benchmarking End-To-End Performance of AI-Based Chip Placement Algorithms (NeurIPS 2025 · similarity 0.86)
- Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer (NeurIPS 2024 · similarity 0.86)
- Retrieval-Guided Reinforcement Learning for Boolean Circuit Minimization (ICLR 2024 · similarity 0.84)
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.