This small notebook should allow students of GU's MLT programme to estimate their CO2 consumption when training models. If you see any mistake, errors, or have comments: please open a PR or send me an email!
Some disclaimers:
- Network storage isn't counted
- When CPU-only jobs, GPUs are considered to draw 0 power
- RAM, HDDs, SSDs are guesstimates
- PSUs are not 100% efficient so there is overhead that needs to be counted too
- I've not read in details the "CO2 to bananas" conversion webpages, nor checked whether these values held for Sweden
Other interesting things:
- Strubell, E., Ganesh, A. and McCallum, A., 2019. Energy and policy considerations for deep learning in NLP. ACL 2019. Link: https://www.aclweb.org/anthology/P19-1355/