Adaptive GPU Power Capping

This repository contains the code, datasets, and results for the paper:

Adaptive GPU Power Capping: Balancing Energy Efficiency, Thermal Control and Performance
Tanish Desai*, Jainam Shah*, Gargi Alavani, Snehanshu Saha, Santonu Sarkar
HPDC ’25, Notre Dame, Indiana, USA

Overview

This project presents a machine learning-based approach to dynamically determine optimal GPU power caps using real-time system metrics. It targets enhanced energy efficiency and thermal control with minimal performance trade-off.

Key Results

• Up to 12.87% energy savings and 11.38% temperature reduction
• Less than 3.3% performance overhead
• Validated on YOLOv8 and BERT workloads
• Outperforms or matches static power capping methods

Repository Structure

adaptive-gpu-power-capping/
│
├── Datasets/               # Raw dataset used for model training
├── models/                 # Trained ML models (CatBoost, XGBoost, etc.)
├── evaluation/             # Scripts and logs for testing on YOLO/BERT
├── results/                # Figures and tables from the paper
├── tree_models.py          # Script to train models
├── merged_results.csv      # Processed dataset used for model training
└── README.md               # Project overview and instructions

Setup & Requirements

Install dependencies:

pip install -r requirements.txt

(Optional) Setup virtual environment:

python -m venv venv
source venv/bin/activate

Running the Code

To train the model:

python src/train_model.py --model catboost

To evaluate on a test workload:

python src/run_inference.py --input data/test_metrics.csv

Dataset

Collected from 7 GPU workloads on NVIDIA RTX 4000 Ada using varying power caps. Metrics include:

• Power, Temperature, Energy
• GPU Utilization, Memory Utilization, Frequency
• Label: Optimal power cap minimizing Energy-Delay Product (EDP)

Citation

If you use this work, please cite our paper:

@inproceedings{desai2025adaptive,
  title={Adaptive GPU Power Capping: Balancing Energy Efficiency, Thermal Control and Performance},
  author={Desai, Tanish and Shah, Jainam and Alavani, Gargi and Saha, Snehanshu and Sarkar, Santonu},
  booktitle={HPDC '25},
  year={2025}
}

Contact

For questions or contributions, contact:
f20220053@goa.bits-pilani.ac.in, f20221182@goa.bits-pilani.ac.in

---