This repository contains datasets for training machine learning models that predict the thermal spatial distribution of chips based on their performance metrics. The data is collected on Intel i5-3337U CPU, Intel i7-8650U CPU, AMD Ryzen 7 4800U CPU, AMD Ryzen 7 7730U CPU, AMD Ryzen AI 5 340, NVIDIA GeForce RTX 4060 GPU, and Google Coral M.2 TPU. We will continue to expand this database going forward. Please stay tuned.
The data is stored in serialized Python object format and can be read using Python's pickle module. Each file contains a Python dictionary with two items. "input" stores a 2D/3D NumPy array representing performance metrics or other inputs for the model prediction. "output" stores a 3D array representing the thermal map data. The first dimension indicates the data point number within the dataset. For each data point, the input can either be a vector (1D) or a time series (2D), while the output is a 2D thermal map (in degrees Celsius).
Due to GitHub's file size limitations, only some samples are stored here. The complete datasets are available upon request.
Lu J, Tan S X, "Thermal Map Dataset for Commercial Multi/Many Core CPU/GPU/TPU", Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD, vol. 33, pp. 1–7, 2024. DOI: 10.1145/3670474.3685963. The paper can be downloaded here: Commercial Thermal Map Dataset.pdf
AI-Empowered Thermal Modeling and Run-Time Management for Manycore Processor and Chiplet Designs
Files starting with CPU_i5 and CPU_i7 contain data for the Intel i5-3337U and i7-8650U CPUs. Each file corresponds to continuous recordings of CPU performance metrics and thermal maps over time under a specific task. For each data point, the input is a vector and the output is a heat map. You can stack the performance metric data from the time points preceding a given time to form a time series for training a time series model.
Intel i5-3337U
| Parameter | Value |
|---|---|
| CPU cores / threads | 2C / 4T |
| Clock speed | 1.8 / 2.7 GHz (base / boost) |
| Process node | 22 nm (Ivy Bridge) |
| Peak performance | ~45 GFLOPS (FP64, est.) |
| TDP | 17 W |
Intel i7-8650U
| Parameter | Value |
|---|---|
| CPU cores / threads | 4C / 8T |
| Clock speed | 1.9 / 4.2 GHz (base / boost) |
| Process node | 14 nm (Kaby Lake-R) |
| Peak performance | ~200 GFLOPS (FP32, est.) |
| TDP | 15 W |
Fig. 1 — Thermal map of Intel i5-3337U
Fig. 2 — Thermal map of Intel i7-8650U
Fig. 3 — i7-8650U with temperatures at sensor and true hot spot
CPU_R7_4800U.pkl contains data for the AMD Ryzen 7 4800U CPU. For each data point, the input is already a time series and the output is the corresponding thermal map.
| Parameter | Value |
|---|---|
| CPU cores / threads | 8C / 16T (Zen 2) |
| Clock speed | 1.8 / 4.2 GHz (base / boost) |
| Process node | 7 nm |
| Peak performance | ~1 TFLOP (CPU FP32, est.) |
| TDP | 15 W |
AMD.Ryzen.7.4800U.HotSpots.Released.mov
The thermal map video above shows the changing hot spots across different cores over time.
ADM.Ryzen.7.4800U.temp.and.power.with.scale.marks.mp4
The thermal map vs. resulting power density (heat flux map) over time, with different hotspots.
CPU_R7_7730U.pkl contains data for the AMD Ryzen 7 7730U CPU (to be added soon). For each data point, the input is already a time series and the output is the corresponding thermal map.
| Parameter | Value |
|---|---|
| CPU cores / threads | 8C / 16T (Zen 3 refresh) |
| Clock speed | 2.0 / 4.5 GHz (base / boost) |
| Process node | 7 nm |
| Peak performance | ~1.2 TFLOPs (CPU FP32, est.) |
| TDP | 15 W |
AMD.Ryazen.7730U.hotspots.mp4
The thermal map video above shows the changing hot spots across different cores over time.
CPU_R_AI5_340.pkl contains data for the AMD Ryzen AI 5 340 (to be added soon). This is a Strix Point SoC featuring a 4-core/8-thread Zen 5 CPU, an integrated GPU, and a dedicated NPU, fabricated on a 4 nm process. Key specifications are listed below.
| Parameter | Value |
|---|---|
| CPU cores / threads | 4C / 8T (Zen 5) |
| Clock speed | ~3.0 / 4.0 GHz (base / boost, est.) |
| Process node | 4 nm (Strix Point) |
| NPU performance | ~50 TOPS |
| iGPU performance | ~5–6 TFLOPs (est.) |
| TDP | 28–45 W (configurable, CPU + GPU + NPU) |
AMD.Ryzen.AI.5.340.thermal.run1.mp4
The thermal map video above shows the changing hot spots across different cores over time for AMD Ryzen AI 5 340.
AMD.Ryzen.AI.5.340.power.run1.mp4
The resulting power density map from the thermal map
GPU_RTX_4060.pkl contains data for the NVIDIA GeForce RTX 4060 GPU. For each data point, the input is already a time series and the output is the corresponding thermal map.
| Parameter | Value |
|---|---|
| CUDA cores | ~3,072 |
| Clock speed | 1.83 / 2.46 GHz (base / boost) |
| Process node | 5 nm (Ada Lovelace) |
| Peak performance | ~15–24 TFLOPs (FP32) |
| TDP | 115 W (desktop) / 35–80 W (laptop) |
Fig. 4 — Thermal map of NVIDIA RTX 4060
| Parameter | Value |
|---|---|
| CPU cores | 8C (4× Kryo 265 Gold / A73 + 4× Kryo 265 Silver / A53) |
| Clock speed | 2.4 GHz (A73) / 1.9 GHz (A53) |
| Process node | 6 nm (TSMC) |
| Peak performance | ~1 TOPS (Hexagon DSP) |
| TDP | ~6–8 W |
The figure below shows the temporal evolution of the thermal map of the Qualcomm SM6225 (Snapdragon 680 4G) SoC, illustrating how on-chip temperature distributions change over time under workload execution.
moto_thermal_short.mp4
This figure compares the initial thermal map with the corresponding derived power-density (heat flux) map over time. Distinct hotspot regions are clearly observed, demonstrating the relationship between localized power dissipation and temperature rise.
thremal.vs.power.map.Snap.Dragon.680.4G.mp4
An enlarged view of the thermal map is shown alongside the corresponding power-density (heat flux) map, providing finer spatial resolution of hotspot formation and evolution over time.
moto_thermal_short_power_dct_comparison_v2.mp4
TPU_Google_Edge.pkl contains data for the Google Coral M.2 TPU. Due to its task-specific nature, we do not use performance metrics to predict real-time temperature. Instead, we use the features of the machine learning tasks deployed on it to predict the steady-state temperature during runtime. For each data point, the input is a feature vector of the workload, and the output is the steady-state thermal map.
| Parameter | Value |
|---|---|
| Architecture | 1× Edge TPU (ASIC) |
| Clock speed | N/A (fixed-function AI accelerator) |
| Process node | 28 nm (GlobalFoundries) |
| Peak performance | 4 TOPS (INT8) |
| TDP | ~2 W |
Fig. 5 — Thermal map of Google Coral M.2 TPU
If you use this dataset in your research or project, please cite it using the following BibTeX entry:
@inproceedings{mlcad2024commercialthermalmapdataset,
author={Jincong Lu and Sheldon X.-D. Tan},
title={Thermal Map Dataset for Commercial Multi/Many Core CPU/GPU/TPU},
booktitle={Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD},
series={MLCAD '24},
year={2024},
location={Salt Lake City, Utah},
url={https://dl.acm.org/doi/10.1145/3670474.3685963},
doi={10.1145/3670474.3685963},
publisher={ACM},
address={New York, NY, USA},
}