Post #3222293
2025-12-16 13:57 UTC
@sarahalang@fedihum.org Thank you, there's a lot of insight in this paper!
A couple of questions, if I may:
1. Did you only consider LLM-generated tokens in your calculations? Or also the tokens in the prompt that need to be processed before generating new ones?
2. Am I right that the total energy consumed by LLMs was around 40 kWh (or 200 kWh, if you consider the 5x pessimistic estimate)? I was expecting this figure to be shown prominently in the paper, but couldn't find it stated directly.
Replies (3)
-
@sarahalang@fedihum.org 2025-12-16 17:31
@osma@sigmoid.social we wrote this a few months ago, so I need a moment to check with the colleagues but will get back to you with the answers!
-
@sarahalang@fedihum.org 2026-01-07 14:58
@osma@sigmoid.social sorry for the belated response! 1) we only used completion, i.e. output, tokens as that is what the tools we used do, too, although I’m not 100% sure how this is justified.
-
@sarahalang@fedihum.org 2026-01-07 15:10
@osma@sigmoid.social and you are right: 2) we didn’t put the total energy numbers anywhere prominently. They are the following: Total energy in kWh (min-max): 37-43 Sum 5x (min-max): 185-215 The basis data for it technically was in C table 5 but not a summary like this.