Nobody ever cares 
I mean, government.
Those those are interesting questions, especially that no one thinks about it. These are things my university has just started discussing for the energy use of our clusters (mostly for processing petabytes of genomics data), but it is a topic of discussion and something the cluster managers are well aware of (Iâve personally discussed it with them).
The answer could end up being that in the broader context of society the energy consumption of scientific models is an irrelevant cost for a significant gain. But it might not be, I have no idea really. I think we should know more about it than we do currently. The concern of this thread (AI) might end up quite a few orders of magnitude larger though, due to its generality. Like Iâm considering using it for filtering massive amounts of historical text to extract environmental data, and probably that will use more energy than anything I do at the moment.
Yeah, we all like to believe this. We use resources like they are given by God and request more every year. The only energy discussion Iâve heard was âhow are we going to keep this thing coolâ.
This is interesting. I would never guess that AI is so much more energy hungry than say GFD models.
Ok yeah thatâs kind of depressing.
Maybe very few AI use cases will match your models individually. But in aggregate Iâm guessing they will easily. And that will be hidden in huge distant cloud warehouses somewhere rather than our current very obvious on-site clusters we have known energy use metrics for.
At least half the people I work with already pay for chat GPT to use for science. Its still not near our cluster budget but its not insignificant, and this thing seems to be only just starting.
Energy use is an interesting discussion. I for one am much more concerned about energy efficiency than CO2 reduction. When is the last time we warmed our car for 5 minutes before getting in and driving off. Something I never do but see others doing this constantly in cold weather.
And every time that we scuttle a pipeline or drilling project for energy in the west, we are essentially saying we want to pay for terrorism as international energy dollars are often used to fund terrorism.
I do see some industries offloading energy use requirements to off peak hours. For a TMP pulp mill this can be 75 MW for one refiner line. The pulp can be stored in tanks for use during peak energy times, or production can be curtailed when energy costs are too high. And money is definitely a driver here.
Another interesting relationship is the use of computing for predictive maintenance. If an incipient fault is not rectified during scheduled downtime, then the plant will be using energy while not producing, a major financial and energy problem.
An interesting topic could be the best use of computing to utilize the earthâs resources in the most effective manner, but that would be another thread.
I do it.
I avoid unnecessary driving anyway, but driving with a fogged windscreen is a risk, and causing a crash is no good. For the environment BTW too.
There is no such thing as a free breakfast. Well, sometimes there is, but mostly no.
1 Like
As this post doesnât seem to go anywhere, perhaps it is time to quote Jean Yanne: âEveryone wants to save the planet, but no one wants to take out the trashâ
4 Likes
If energy consumption with Julia is of interest to you, check out: I Have a dream! a Green dream -- Does Julia save energy? - #14 by garborg
2 Likes
That is it. If some computation is really expensive in terms of energy use, in most case you will be presented the bill, directly or indirectly.
If you donât see the bill, then it is probably more or less peanuts.
I disagree. Just those who claim they want to save the planet and those who take out the trash are mostly not the same people in my experience.
1 Like
Relax, it was a joke from a great comedianâŚ
2 Likes
I couldnât agree more. My experience (watching the energy sector) is you have people who know, people who act, and people who talk (loud). The last group doesnât intersect much with the first two.
That being said, I think this thread would be most useful and productive if the discussion was kept at the actual energy use of AI. When I see timings per token, and 25 seconds of CPU time (on a consumer grade machine) to produce a paragraph of text, I wonder how this can possibly make any economic sense outside of the intellectual curiosity.
1 Like
Thatâs not a problem with AI or ML, thatâs a problem with the journals review process. Maybe that is where your actual gripe lies.
1 Like
The journal editors and review processes are most definitely at fault in computational mechanics
Iâm sure thereâs plenty of blame to go around among the editors and reviewers. Papers with AI/ML mentions have a lot of sex appeal at the funding agencies ATM.
And letâs admit it, these people rarely know of the energy footprint of AI/ML or donât care very much. Believe me, I have tried to talk sense into some of my colleagues. Mixed success.
But it seems to me that awareness is lacking among the (scientific) programmer/ software developer as well. They know the models take a long time to train, but the connection with energy and economics of the enterprise is not necessarily clear in their minds.
1 Like
But then, how high is it, in terms of kWh per solved problem of published paper. Do you have any estimation?
There was a paper I recently reviewed where a model was described that was trained on a 4kW cluster (4 machines) for a little more than day. So perhaps 80-90 kWh? Now this model was supposed to be used for model reduction to make predictions of brittle fracture. The reduced model would run for a few minutes as opposed to an hour for the full model on a single machine. So let us say the reduced model saves an hour of computing. It was expected to be run a few dozens of times. This barely amortizes the initial outlay of training the model (actually, more likely it does not).
An anecdote, no more. It would be good to have statistics, for sure.
1 Like
80-90 kWh cost less than 40$ even in Germany.
In experimental science, the spending on materials only would be typically measured in thousands $$ per paper, a large part of it being (hidden) energy costs. Scientific progress costs money. I think 50$ for a result or paper is not of any real significance.
Surely I canât judge whether AI/ML makes sense in this specific case.
I found this article by Cory Doctorow to be an interesting take on the subject: What Kind of Bubble is AI?.
Do the potential paying customers for these large models add up to enough money to keep the servers on? Thatâs the 13 trillion dollar question, and the answer is the difference between WorldCom and Enron, or dotcoms and cryptocurrency. [âŚ] Just take one step back and look at the hype through this lens. All the big, exciting uses for AI are either low-dollar (helping kids cheat on their homework, generating stock art for bottom-feeding publications) or high-stakes and fault-intolerant (self-driving cars, radiology, hiring, etc.).
9 Likes
This topic was automatically closed after 2 days. New replies are no longer allowed.