Research map / Vision-language and multimodal / Cross-modal learning
FlexPlanner: Flexible 3D Floorplanning via Deep Reinforcement Learning in Hybrid Action Space with Multi-Modality Representation
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 FlexPlanner, an RL method with hybrid action space (position, layer, aspect ratio) and multi-modal state (vision, graph, sequence) for 3D floorplanning that avoids heuristic representations and handles cross-die alignment.
Problem
3D IC floorplanning must place blocks across stacked dies while handling flexible aspect ratios and cross-die alignment; heuristic representations limit the solution space and cannot easily enforce alignment, and prior learning-based methods are 2D-only.
Approach
A learning-based floorplanner that directly outputs block positions, die layer, and aspect ratio (hybrid action space) from vision, graph, and sequence modalities, with alignment and position masks filtering invalid positions, an asynchronous layer-decision module using a Transformer over per-die queues, and a reward combining local step advantages with a global episode baseline.
Key results
- Alignment score improves from 0.474 (MaskPlace, second best) to 0.940 averaged across circuits.
- Average HPWL reduced to 249,053 vs 296,791 for the second-best method, about 16% reduction.
- Zero-shot transfer from n100 to other circuits yields alignment of 0.859 to 0.970 across tested circuits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FlexPlanner instead of MaskPlace | 3D floorplanning · MCNC and GSRC (Avg.) | alignment score: 0.474 -> 0.940 | Helped |
| use FlexPlanner instead of Wiremask-BBO | 3D floorplanning · MCNC and GSRC (Avg.) | HPWL: 296,791 -> 249,053 | Helped |
| add asynchronous layer decision instead of synchronous die-by-die placing (ablation) | FlexPlanner · n100 | alignment score: 0.850±0.042 -> 0.961±0.017 | Helped |
| add alignment mask (ablation) | FlexPlanner · n100 | alignment score: 0.349±0.026 -> 0.961±0.017 | Helped |
| add graph modality (ablation) | FlexPlanner · n100 | alignment score: 0.874±0.027 -> 0.961±0.017 | Helped |
| add sequence modality (ablation) | FlexPlanner · n100 | alignment score: 0.860±0.021 -> 0.961±0.017 | Helped |
| use local-advantage and global-baseline reward instead of sparse GraphPlace reward (ablation) | FlexPlanner · n100 | alignment score: 0.840±0.018 -> 0.961±0.017 | Helped |
| use local-advantage and global-baseline reward instead of MaskPlace difference reward (ablation) | FlexPlanner · n100 | alignment score: 0.721±0.054 -> 0.961±0.017 | Helped |
Surprising findings
- Synchronous die-by-die placing degrades to poor layouts when the layer decision is not guided by the global baseline reward, a collapse that occurs in MaskPlace-style dense rewards.
- Removing the alignment mask drops alignment score from 0.961 to 0.349 on n100.
Limitations
- Only evaluated with two dies and a fixed grid size of 128; thermal and other 3D constraints not addressed.
- Baselines were extended by the authors to 3D and may not reflect their original tuned performance.
- Alignment minimum set to alpha=1.0 (most challenging setting), results under other settings not explored.
Experimental setup
Datasets: MCNC GSRC
Benchmarks: MCNC GSRC
Metrics: alignment score HPWL overlap out-of-bound runtime
Baselines: 3D-B*-SA RL-CBL Wiremask-BBO GraphPlace DeepPlace MaskPlace
Builds on: PPO MaskPlace GraphPlace DeepPlace Generalized Advantage Estimation
Closest papers by meaning
- DeepLayout: Learning Neural Representations of Circuit Placement Layout (ICML 2025 · similarity 0.86)
- See it to Place it: Evolving Macro Placements with Vision Language Models (NeurIPS 2026 · similarity 0.86)
- CircuitFusion: Multimodal Circuit Representation Learning for Agile Chip Design (ICLR 2025 · similarity 0.84)
- Chip Placement with Diffusion Models (ICML 2025 · similarity 0.83)
- Reinforcement Learning within Tree Search for Fast Macro Placement (ICML 2024 · similarity 0.83)
- DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale (ICLR 2025 · similarity 0.82)
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.