Hi all,
I’m an independent researcher (B.Sc. CSE, Rangamati Science and Technology University, Bangladesh) and just finished a literature review on a question that seemed under-examined: does model compression (quantization, pruning, distillation) actually reduce measured energy, or just FLOPs?
Drawing on a corpus of 53 works (2019-2026), the review finds a few things that surprised me. Software energy estimators and hardware counters (RAPL, NVML) disagree by margins comparable to the savings compression claims to deliver. FLOPs correlate weakly with measured energy once memory bandwidth, batch size, and backend kernel realization are accounted for. And compression rankings on one hardware platform don’t reliably transfer to another.
Zenodo (DOI): Energy Measurement of Compressed Deep Learning Models: A Literature Review | Zenodo Data/tables repo: GitHub - ShohanRony/energy-measurement-compressed-dl-review: Literature review: energy measurement of compressed deep learning models · GitHub
Separately, I’d appreciate any pointers from this community. I’m a first-time arXiv submitter and need a cs.LG endorsement to post this (and a related Master’s research proposal) there. If you’re an active cs.LG author on arXiv and this area is anywhere near your work, I’d be glad to share more detail, feel free to reply or DM.
Happy to hear feedback on the review itself too, especially if I’ve missed relevant work.
Thanks for reading!