day3/thermal-l1-staging

IC-ThermBench: S1 source-suite datasets

Scope S1 of IC-ThermBench collects established within-design thermal tasks from the ARO → Therm-FM research line: the Alpha EV6 HotSpot packages (single-, quad-, and octa-core at two grid refinements) and two industrial packages. The tasks keep their original definitions, grids, and fidelity settings; they are collected, not regenerated.

• IC-ThermBench paper: https://arxiv.org/abs/2608.23977
• Code and evaluation: https://github.com/Day333/ThermalBench
• S2–S5 datasets: https://huggingface.co/datasets/day3/IC-ThermBench-S2-S5

Files

File	Contents
thermal_steady.tar.gz	8 steady-state datasets (≈4.5 GB compressed, ≈5.9 GB unpacked)
thermal_transient.tar.gz.part00–part03	3 transient datasets, split into four parts (≈17.6 GB compressed, ≈131 GB unpacked)

Reassemble and unpack:

tar -xzf thermal_steady.tar.gz
cat thermal_transient.tar.gz.part0* > thermal_transient.tar.gz   # Windows: copy /b thermal_transient.tar.gz.part0* thermal_transient.tar.gz
tar -xzf thermal_transient.tar.gz

Each dataset folder holds input.mat and output.mat (MATLAB v7.3 / HDF5, key data). Temperatures are in kelvin. Each archive also contains a README.md describing the format.

Steady-state datasets

input.mat is (N, P, L, H, W) and output.mat is (N, L, H, W), where P = 4 channels are power density, x, y, and z/layer index, and L is the number of layers.

Folder	input.mat	output.mat
HS_SC_refine1	(5000, 4, 2, 55, 55)	(5000, 2, 55, 55)
HS_SC_refine2	(5000, 4, 2, 87, 87)	(5000, 2, 87, 87)
HS_QC_refine1	(5000, 4, 2, 39, 39)	(5000, 2, 39, 39)
HS_QC_refine2	(5000, 4, 2, 63, 63)	(5000, 2, 63, 63)
HS_OC_refine1	(5000, 4, 3, 85, 85)	(5000, 3, 85, 85)
HS_OC_refine2	(5000, 4, 3, 151, 151)	(5000, 3, 151, 151)
IND_8C	(1000, 4, 2, 101, 101)	(1000, 2, 101, 101)
IND_32C	(1000, 4, 2, 101, 101)	(1000, 2, 101, 101)

Transient datasets

input.mat is (N, P, T, L, H, W) and output.mat is (N, T, L, H, W), where P = 5 channels are power density, x, y, z/layer index, and time-step index.

Folder	input.mat	output.mat
HS_SC_refine2	(5000, 5, 9, 2, 87, 87)	(5000, 9, 2, 87, 87)
HS_QC_refine2	(5000, 5, 5, 2, 63, 63)	(5000, 5, 2, 63, 63)
HS_OC_refine2	(3000, 5, 9, 3, 151, 151)	(3000, 9, 3, 151, 151)

License and attribution

The S1 source tasks retain their upstream licenses and citation terms (see LICENSE-DATA). If you use these datasets, please cite ARO and Therm-FM as the source of the tasks, and IC-ThermBench if you use them as part of the benchmark:

@article{icthermbench2026,
  title   = {IC-ThermBench: An Open, Progressive Benchmark for Generalizable
             2.5D/3D-IC Thermal Learning},
  author  = {David Hang and Wenkai Yang and Kuiye Ding and Haiyang Xin and Jacky Wei},
  journal = {arXiv preprint arXiv:2608.23977},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.23977}
}