Flash-CNNCap: Capacitance Extraction via Image Mapping

Accepted at ICCAD 2026. Read the paper on arXiv.

Flash-CNNCap predicts spatial capacitance-contribution maps and reduces them over conductor masks, lowering full-matrix reconstruction from $O(n^2)$ to $O(n)$ model evaluations.

This artifact contains the paper's binary-occupancy training code, exact 13-model ablation, GPU DEF-to-SPEF pipeline, and 14 reported D4_D_k5 checkpoints. Post-paper experiments are intentionally excluded.

Quick Start

The reported runs used the NVIDIA PyTorch 26.02 container and CapBench revision 33939cb5002575c06f19c237e2eae6105e331634:

python -m pip install -r requirements.txt
git clone https://github.com/THU-numbda/CapBench.git
git -C CapBench checkout 33939cb5002575c06f19c237e2eae6105e331634
python -m pip install -e './CapBench[all]'
python -m capbench datasets install nangate45/small

Pretrained Models

Download the optimizer-free paper checkpoints from the v1.0.0 release:

gh release download v1.0.0 \
  --repo THU-numbda/flash-cnncap \
  --pattern 'flash-cnncap-d4-d-k5-*.pth' \
  --dir models

The model-to-dataset mapping and checksums are in models/.

Reproduce the Paper Results

python scripts/ablations.py list
python scripts/ablations.py run --dataset nangate45/small --num-gpus 8 --epochs 100 --repeats 5
python scripts/best_model_eval.py run --model D4_D_k5 --gpu-ids 0,1,2,3,4,5,6,7 --epochs 100

See docs/reproduction.md for the complete protocol and full-pipeline/README.md for deployment. The optional HDF5 cache stores binary occupancy only and does not change the paper's input encoding.

Citation

If you use Flash-CNNCap in academic work, please cite both the method paper and the CapBench dataset used to train and evaluate it. Machine-readable metadata is also available in CITATION.cff.

Flash-CNNCap

Flash-CNNCap: Capacitance Extraction via Image Mapping, ICCAD 2026:

@inproceedings{rodriguez2026flashcnncap,
  title         = {Flash-CNNCap: Capacitance Extraction via Image Mapping},
  author        = {Rodriguez, Hector R. and Huang, Jiechen and Yu, Wenjian},
  booktitle     = {2026 IEEE/ACM International Conference on Computer-Aided Design (ICCAD)},
  year          = {2026},
  eprint        = {2607.23877},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2607.23877}
}

CapBench

CapBench: A Multi-PDK Dataset for Machine-Learning-Based Post-Layout Capacitance Extraction, DAC 2026:

@inproceedings{rodriguez2026capbench,
  title         = {CapBench: A Multi-PDK Dataset for Machine-Learning-Based Post-Layout Capacitance Extraction},
  author        = {Rodriguez, Hector R. and Huang, Jiechen and Yu, Wenjian},
  booktitle     = {Proceedings of the 63rd ACM/IEEE Design Automation Conference (DAC)},
  year          = {2026},
  eprint        = {2604.11202},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AR},
  url           = {https://arxiv.org/abs/2604.11202}
}

See THIRD_PARTY_NOTICES.md for technology-data attribution.