Sustainable AI Data Compression Case Study: Profiling BOA on BEDE
Project Overview
During my 8-week internship, I developed a reproducible workflow to measure the environmental impact of training and running a deep learning model for data compression i.e BOA Constrictor, a lossless data compression method for high-energy physics data that requires large-scale machine learning training on a high-performance computing system.
My main goal was to quantify the energy consumption and carbon emissions associated with training and using the compression method, and to turn this work into a practical training resource. I was able to run the workflow on 3 GPUs i.e Bede, RTX5090 and Tesla T4 and compared the carbon cost of BOA Constrictor against classical compressors such as LZMA and ZSTD.
What were the key results of your research project?
- No savings at scale because energy grew in step with input data size.
- The choice of GPU changed speed far more than energy use.
- Since BOA beats standard tools in compression ratio, neural compression is only worth its carbon cost when the data will be stored for a long time.
- Also created reproducible benchmark scripts and a CodeCarbon learning guide.
GitHub Repository: https://github.com/HanzilaH/N8CiR-SusTraIN-data-compression-notebooks/tree/main
A presentation of this research will be shared here and on our YouTube site when available.
How do you feel you have benefitted from completing this internship and has it made you consider future career paths?
This internship gave me hands-on experience conducting real-world research and working alongside exceptional PhD researchers. I learned to design fair benchmarks, measure energy use with CodeCarbon, and work with GPUs and large datasets. I also improved my Python, got better at making code reproducible, and presented my results to a research audience.
It made me think about the environmental cost of computing, which I hadn't considered before. It has made me interested in sustainable computing as a future career
Download presentation slides