The heavy computing involved in AI - training and running deep learning algorithms, with big datasets - consumes power and resources. How much should we be concerned about its environmental impact?
In 2019, the paper Energy and Policy Considerations for Deep Learning in NLP made quite a big splash by explicitly calculating the CO2 footprint of some example deep learning algorithms. It led to headlines like this:
"Training a single AI model can emit as much carbon as five cars in their lifetimes" (MIT Technology Review)
If that sounds worrying: well, take a moment and think whether that headline seems credible. If training an AI model takes as much power as 5 cars over their whole lifetime, how come my home laptop can do it in a couple of hours or days? -- The calculation they're referring to is not for an "average" AI model, but for one of the biggest models ever (at the time), and also for the fact that they're training it many times over, during an optimisation procedure.
Still, the paper makes a very good initial analysis, including some smaller models as well as the bigger ones. I found it interesting that they analysed their own logfiles in their lab, to determine that in their previous work they had executed a training run almost 5000 times!
A 2022 paper from Google re-analyses the claims and gives a lot of numerical detail from Google's own datacenters, giving a dramatic and reassuring claim:
The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink
However, the headline claim that AI's footprint "will shrink" is an absolute over-claim, based on little more than some optimistic guesses in the paper's discussion section. The authors really should have thought twice about choosing that title for the paper, since there's a very big rhetorical impact if Google announces that AI's carbon footprint is maybe nothing to worry about.
They do some very good work on carbon emissions, including running their own low-carbon power generation from windfarms. Their own analysis found that one of their datacentres in Oklahoma is 92% carbon-free (due to good conditions for wind energy generation), and as a result they moved the location for their big AI training to there. The claim is thus that if we all did things like this, we - just like Google - can maintain our impressive rate of progress while our carbon footprint plateaus and decreases.
So what's wrong with Google's claim?
Well firstly, we can't all move our AI servers to Oklahoma. Even if we could, many wouldn't. If we decarbonise our energy sector entirely, then of course AI's footprint will shrink, but - well, you've heard the debates about that raging for years.
Secondly, the claim is that better hardware is more efficient - such as Google's new generation of TPU chips, discussed in the paper. This is true, but again: not everyone can or would move to these. The older GPU and TPU hardware is unlikely to go into landfill: it'll continue to be used, for AI or for other computation.
And thirdly there's the famous Jevons paradox, not at all addressed in the paper: if AI becomes more efficient, this will certainly tempt people to do more of it. This has clearly happened already, with companies training AI models so big that they would have been inconceivable ten years ago -- many billions of parameters in one model, and many GPU-years of power used to train them.
The ignorance of Jevons paradox is clearest when the paper dismisses the idea that AI inside mobile phones might be a big deal (their global footprint is much smaller than the footprint of the big data centres). But it's a pretty safe bet that, as TPUs in mobile phones get more and more usage by software developers, and mobile apps powered by on-device AI find ever more diverse uses and popularity, they'll become a bigger and bigger part of the footprint of smartphone usage. Conspicuously, in the paper the authors project various changes in AI datacentres into the future, but they don't bother to make any projections about smartphones or other edge devices.
A recent opinion piece in the Communications of the ACM (Our house is on fire: The climate emergency and computing's responsibility) puts it all very clearly in the big picture:
"Computing’s footprint has risen steadily despite becoming much more efficient in the transmission and storage of data over the last 50 years. Greater efficiency leads, almost always, to growth in overall carbon emissions, as efficiency gains are quickly swamped by the desire to do more; hence we see global emissions increasing decade after decade despite continual efficiency gains in every sector.
"However, efficiencies delivered by computing technology could play a vital role in enabling continued functionality within a resource-constrained future. The computing industry should be lobbying strongly for the introduction of a carbon constraint!"