Vibeleaderboard
← All Intel
Intel / article

JouleShare: Request-Level Energy Attribution for Batched LLM Serving

Source
arxiv.org
Author
Qi Luo, Kunlin Li, Ziwen Wang, Dongsheng Wang, Yun Chen
Date
Why it matters

If you bill, chargeback or report energy per request, the -proportional rule most systems use is substantially wrong under batching. This gives measured Shapley ground truth and a cheap feature-based calibrator that closes most of the gap.

Terms in this piece · Glossary
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Recommended reads
Comments

Checking sign-in…

Loading comments…