Considering the title is already clearly wrong since I can generate many images, on device, even while in airplane mode, I don’t have high hopes for the study.

The underlying calculations apparently depend on the following CSV: https://github.com/mlco2/codecarbon/blob/master/codecarbon/d...

Two problems I see here

1. They are using AWS, which is not in the data set

2. The calculation set seems questionable to me. The GCP Iowa data center has some of the best renewable usage, are they using the data center total (a larger data center with more renewable can still have more total emmissions vs only 8% of the compute I use should be included in the calculation. Either way, the underlying data for the calculation is out of date

edit...

I opened an issue with the codecarbon project to see if we can get more clarity on things. I've include links there for the breadcrumbs I followed

https://github.com/mlco2/codecarbon/issues/483

My phone's battery is roughly 50KJ. A single nvidia a100 is spec'd at 250W max.

That's 3.3 minutes to consume 50KJ.

But that only considers a single GPU, the rest of the system will be consuming a good chunk of power. You also have to consider the entire datacenter's energy use.

I'd say the claim is within an order of magnitude of being correct, but I seriously doubt it.

ah, I expected article talks about this is too cheap or AI eat our lunch, but it is about carbon.