clanoman/hotspot-chiplet-floorplan-thermal

HotSpot Chiplet Floorplan Thermal Dataset

Dataset Description

The HotSpot Chiplet Floorplan Thermal Dataset contains 10,000 synthetic steady-state examples for heterogeneous chiplet floorplans. Each combines a fixed structural family, a per-chiplet power/activity workload, the fixed v2.0 thermal configuration, and a HotSpot-generated 64×64 Layer-0 temperature field in kelvin.

ChipTherm is the research project that generated and uses the dataset. The release is model-independent: raw geometry and per-chiplet power let researchers construct their own representations.

This is simulated, not measured-silicon, data and is not a validated model of commercial packages.

Dataset Structure

The dataset contains 50 immutable structural families and 200 workloads per family. The numerical payload in data/ consists of:

• —metadata.parquet, containing scalar metadata and shard/row locations; and
• —20 Zstandard-compressed files named samples-00000-of-00020.parquet through samples-00019-of-00020.parquet, with exactly 500 samples per shard.

Parquet packaging avoids 10,000 separate .npy labels. Every shard row contains model-independent layout information, per-chiplet power, metadata, and temperature_K: 4,096 row-major float32 values representing a 64×64 field in kelvin. metadata.parquet omits the field and raw nested inputs.

Data Fields

Important fields include:

• —sample_uid, family_uid, workload_uid
• —family_category, placement_style, chiplet_count
• —package_width_mm, package_height_mm
• —layout_json: package footprint and chiplet names, types, positions, and dimensions
• —chiplet_power_W_json, active_chiplet_names, total_power_W
• —active_chiplet_fraction, dominant_chiplet_share
• —workload_stratum, workload_cell, power_regime, topology_regime
• —both frozen evaluation-protocol memberships
• —structural, workload, and temperature-content SHA-256 identifiers
• —temperature_K, with shape, layer, units, and orientation metadata

Lengths are in millimetres, power is in watts, power density is in W/mm² where present, and temperature is in kelvin.

Dataset Creation

Family definition

A family fixes its package footprint, chiplet composition/types/geometry, thermal configuration, and HotSpot settings. Only power and activity vary among its 200 samples. Geometry was generated deterministically from the frozen v2 specification with seed 20260721.

Workload generation

The 200 workloads form a frozen 10-power-regime × 20-activity-topology design covering balanced, type-specific, compute/memory-dominant, single-source, clustered, near/far, distributed, edge/corner, central, symmetric, and asymmetric activity. Per-chiplet powers and content hashes are retained.

HotSpot provenance

Targets are parsed from Layer 0 of HotSpot's steady grid output. HotSpot uses a 64×64 grid with average grid-to-block mapping; leakage, detailed-package modeling, and the secondary thermal path are disabled. The fixed v2.0 stack, boundaries, and settings are recorded under configs/.

HotSpot software is not included. manifests/provenance.json records the executable SHA-256 and an observed source checkout, but they are not cryptographically bound.

Please cite the relevant HotSpot work:

• —K. Skadron et al., “Temperature-Aware Microarchitecture,” ISCA 2003.
• —W. Huang et al., “Compact Thermal Modeling for Temperature-Aware Design,” DAC 2004.
• —R. Zhang, M. R. Stan, and K. Skadron, “HotSpot 6.0: Validation, Acceleration and Extension,” University of Virginia Technical Report CS-2015-04.

Evaluation Protocols

Seen-family workload interpolation

Evaluate new workloads on the 40 training-represented families: 6,400 training, 800 validation, and 800 test samples. This measures workload interpolation, not package-family generalization.

Unseen-family structural generalization

Evaluate complete families absent from training using a family-disjoint 40/5/5 partition. All workloads inherit their family's assignment, producing 8,000 training, 1,000 validation, and 1,000 test samples.

Rotational folds are not part of this release because authoritative validation does not establish them as completed protocols.

Recommended Metrics

Report full-field MAE, global cell-wise RMSE, peak-temperature absolute error, and unweighted per-family aggregates in kelvin. For unseen-family structural generalization, also report mean, median, and worst-family results. Fit normalization only on the applicable training partition and always name the protocol.

Limitations

• —All data are synthetic and inherit HotSpot's assumptions and numerical limitations.
• —There are no measured silicon data or validated commercial-package models.
• —Thermal stack, ambient convention, cooling configuration, grid resolution, and HotSpot settings are fixed in v2.0; material/cooling generalization is not evaluated.
• —Detailed package and secondary thermal paths are disabled.
• —Rectangular floorplans and a fixed 64×64 grid simplify real packages; the taxonomy is designed coverage, not commercial-product sampling.

Licensing

The dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0); see LICENSE-DATA. It covers project-generated geometry, workloads, metadata, splits, and HotSpot-generated numerical targets.

CC BY 4.0 does not license or relicense HotSpot. HotSpot software remains governed by its University of Virginia license and is not redistributed here. The ChipTherm software repository is separately licensed under Apache-2.0.

Citation

If you use this dataset, please cite the dataset and the relevant HotSpot publications listed above. Formal ChipTherm publication metadata will be added when available; no paper title, DOI, venue, or additional authorship is asserted here.