Chip Power Prediction with GNNs

TLDR

• Transformed raw chip netlists into PyTorch Geometric graph representations for power prediction across ~50-100 designs.

• Built graph featurization pipelines with structural features and attention layers, then benchmarked linear regression, random forests, gradient boosting, and GNNs.

• Evaluated models on MAE, RMSE, and MAPE, while also benchmarking inference speed and scalability across design complexity.

• Used SHAP analysis to identify which netlist features were driving power estimates.