[hdkrgr]

I do think this will be a useful metric, and it seems obvious that the hyperscalers will have a feature helping you keep track of energy use and emissions of the resources you rented. But why demand this on the level of an individual model/product? For these foundation models, I think it's reasonable to assume they will all be trained on hyperscaler-provided gpu-clusters, so there'll likely be an off-the-shelf funcitonality by AWS/Azure/GCP to report this number, but the draft of the EU AI Act also demands tracking energy use for other 'high-risk' AI systems which companies may plausibly train and/or deploy on-prem. Good luck tracking the per-token energy use of your model that's running on some on-prem server on last-gen GPUs.

[Dylan16807]

Especially for a server GPU, looking up watts and multiplying by time per token should give you a pretty good number.

[hdkrgr]

Sure... but maybe the GPU is sitting idle 40% of the time while still consuming 200W. Should I have to break this idle energy consumption down onto actual use (assuming the server/gpu is only used for this one model)? I guess it would make sense, but... WHO should do this and then continually update the model documentation when idle rates or the hardware changes?

[senko]

The organizations that release the models already provide (brag about) their model performance. They could simply include in the same report the info about the energy spent doing the training/finetuning/inference, per X tokens.

This doesn't necessarily measure every use, just "manufacturer's spec", the same you get for eg energy class for house appliances (at least in the EU). Nobody goes around measuring refrigerator power usage, but when you're buying one, you get a rough indication of how "green" (or not) it is.

[hdkrgr]

I agree, that seems reasonable!

I was referring more to the users of such a system (what the AI Act would call a 'deployer'). They may have significantly less expertise but could still be required to track real-time energy use. Of course, simply referring to the 'energy label' by the provider could be a viable solution.

[Dylan16807]

Listing it per server design (with groups) makes sense to me.

It wouldn't make sense to include measured idle time in the energy numbers you'd include in model documentation. Maybe that could go in a monthly report somewhere, but that's a different topic.