Research map / Graph learning / Knowledge and temporal graphs
PreRoutGNN for Timing Prediction with Order Preserving Partition: Global Circuit Pre-training, Local Delay Learning and Attentional Cell Modeling
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 PreRoutGNN, a two-stage pre-routing timing (slack, slew, delay) predictor combining global circuit auto-encoder pre-training, residual local delay learning with topological-order message passing, multi-head joint attention for cell LUT modeling, and order-preserving graph partitioning.
Problem
Pre-routing slack prediction on large industrial circuits suffers from signal decay and error accumulation along long timing paths, and full-graph GNN training has prohibitive peak GPU memory.
Approach
Combine a self-supervised graph auto-encoder that gives global graph embeddings with a GCN that updates node features in topological-level order using residual message passing (modeling net and cell delay as increments) and an attention layer that mimics LUT query-index-interpolation, plus a partition scheme that pads sub-graphs with neighboring topological levels to keep dependencies.
Key results
- Achieves average un-flatten R^2 of 0.933 on the 7 test circuits for slack prediction, versus 0.589 for TimingGCN (the previous best).
- Ablation: removing global pre-training drops test R^2 from 0.933 to 0.720; removing residual local learning drops it to 0.772.
- Adding the pre-trained encoder to TimingGCN raises its test R^2 from 0.589 to 0.762.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PreRoutGNN instead of TimingGCN | Pre-routing slack prediction · 21 circuits, test average | slack R²uf: 0.589 -> 0.933 | Helped |
| Use PreRoutGNN instead of TimingGCN | Pre-routing timing prediction · 21 circuits, test average | slew R²uf: 0.833 -> 0.934 | Helped |
| Use PreRoutGNN instead of TimingGCN | Pre-routing timing prediction · 21 circuits, test average | net delay R²uf: 0.946 -> 0.967 | Helped |
| Use PreRoutGNN instead of TimingGCN | Pre-routing timing prediction · 21 circuits, test average | cell delay R²uf: 0.706 -> 0.906 | Helped |
| Add a pre-trained graph encoder | TimingGCN · 21 circuits, test average | slack R²uf: 0.589 -> 0.762 | Helped |
| Add global pre-trained auto-encoder (ablation) | PreRoutGNN · 21 circuits, test average | slack R²uf: 0.720 -> 0.933 | Helped |
| Add residual local learning (ablation) | PreRoutGNN · 21 circuits, test average | slack R²uf: 0.772 -> 0.933 | Helped |
| Add multi-head joint attention (ablation) | PreRoutGNN · 21 circuits, test average | slack R²uf: 0.873 -> 0.933 | Helped |
| Fine-tune rather than freeze the pre-trained encoder (ablation) | PreRoutGNN · 21 circuits, test average | slack R²uf: 0.908 -> 0.933 | Helped |
Surprising findings
- Flatten R^2 (pooling four EL/RF corners) gives a much higher score for the prior SOTA (0.866 vs 0.576 un-flatten), so the authors argue flatten R^2 is an inappropriate metric.
- Without fine-tuning the pre-trained encoder, performance drops only slightly (0.908 vs 0.933 test R^2).
Limitations
- Evaluated only on 21 OpenROAD circuits, with slack prediction focused on arrival time as the main target.
- Relies on the RAT being given in external files rather than predicted.
Experimental setup
Datasets: OpenROAD 21-circuit timing benchmark
Benchmarks: OpenROAD 21-circuit timing benchmark
Metrics: R^2 un-flatten R^2 flatten MSE
Baselines: GCNII GAT GINE TimingGCN CircuitGNN Net^2 LHNN
Builds on: TimingGCN GCN NeRF positional encoding
Closest papers by meaning
- NeuralSteiner: Learning Steiner Tree for Overflow-avoiding Global Routing in Chip Design (NeurIPS 2024 · similarity 0.85)
- Bridging Layout and RTL: Knowledge Distillation based Timing Prediction (ICML 2025 · similarity 0.84)
- CircuitFusion: Multimodal Circuit Representation Learning for Agile Chip Design (ICLR 2025 · similarity 0.84)
- DeepLayout: Learning Neural Representations of Circuit Placement Layout (ICML 2025 · similarity 0.84)
- DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale (ICLR 2025 · similarity 0.83)
- DSBRouter: End-to-end Global Routing via Diffusion Schr\"{o}dinger Bridge (ICML 2025 · similarity 0.83)
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.