Graph neural networks (GNNs) are increasingly employed in physical design tasks such as congestion prediction and wirelength estimation. However, their progress is hampered by inconsistent circuit representations and a lack of controlled evaluation protocols.
To address these limitations, researchers have introduced R2G (RTL-to-GDSII), a multi-view circuit-graph benchmark suite. R2G standardizes five stage-aware views with information parity – where each view encodes the same attribute set, differing only in feature attachment points – across 30 open-source IP cores, accommodating up to $10^6$ nodes and edges.
The R2G suite offers an end-to-end DEF-to-graph pipeline covering synthesis, placement, and routing stages. It also includes loaders, unified data splits, domain-specific metrics, and reproducible baselines. By effectively decoupling representation choice from model choice, R2G isolates a confounding factor that previous EDA and graph-ML benchmarks often leave uncontrolled.
Systematic studies conducted with GINE, GAT, and ResGatedGCN models revealed several key insights:
- View choice profoundly impacts performance, dominating model choice, with Test R$^2$ varying by over 0.3 across different representations for a fixed GNN.
- Node-centric views demonstrate superior generalization capabilities across both placement and routing stages.
- Decoder-head depth (specifically 3-4 layers) emerged as the primary driver for accuracy, transforming divergent training outcomes into near-perfect predictions (R$^2$$>$0.99).
The code and datasets for this benchmark are publicly available, promising to accelerate the development and application of GNNs in physical design automation.