What country? Sir Lanka? This isn’t a useful comparison as is, I’ll see if I can dig up actual numbers.
Following for the results of your work here so I can use it in the future.
From this 2023 paper, looks like if all Nvidia AI servers are running 24/7, you’d get an energy consumption of about 5.7–8.9 TWh per year. Nvidia servers make up 95% of the AI market (according to the paper) so that’d be pretty close to what AI servers would consume.
The paper also estimates about 20% of crypto mining GPUs no longer mining etherium converted to AI, which contributed another 16.1 TWh per year.
This doesn’t include some AI, but it should be the majority.
Between those two sources, that gives 23.4 TWh per year. That gives 0.08 exta joules per year per this converter. That’s 22% of Sri Lanka’s energy consumption (which is the lowest country).
So AI in a year uses at much energy as Sri Lanka uses in 3 months. At least in 2023. I’ll see if I can find a more recent study.
So that assumes AI requests use 100 percent of the hardware 100 percent of the time.
Yes, but those servers are pretty ai specific, so that’s a decent assumption. Looks like Nvidia is drastically ramping up production of these servers, so current electricity use might be about 10x, I’m working on it.
100% utilization 100% of the time? That seems like an unlikely figure, right?
100% utilization yes, but server uptimes are in the 99 percent range.
Idle consumption vs full utilization consumption are two very different things, though - this will have a large impact on the final number.
This is the kind of comment I love on Lemmy.
There’s plenty of countries missing from that rankings list, and I bet those are the ones using less energy. Especially considering microstates like Vatican, the statement could be technically correct
I can’t find any info on Vatican City’s energy use, but possibly. You could go even further and compare to not widly recognized countries like sealand, where you have the energy consumption of a residential house or two. But that would be wildly misleading.
Omg it’s the guy from the meme
If you want to be snarky, at least be accurate.
It’s Sir Lanka, disprover of bullshit
This does misunderstand what actually costs the energy – it’s training the models that’s costly, not using the already trained ones. Although to be fair using them increases incentive for new ones to be trained… But yeah asking ChatGPT for a recipe idea isn’t burning an ounce of gasoline.
Using them is also very energy intensive (though much less so than training).
Playing a AAA game amount of energy vs running an entire data center on full blast amount of energy, is the comparison I like to make.
While the order of magnitude is correct, running the bigger models is closer to playing a AAA game on 8 computers at the same time.
Yeah I did forget to consider that a lot of the web hosted models have a whole array of “experts” - Sub-LLMs that help fill in specialized information that a more generalized LLM wouldn’t have. Not a problem for someone running an AI model on their home computer but something that likely happens most times that you’re querying an AI online.
That’s also true, though it’s important to remember that the “experts” aren’t experts in the classical sense. Say you have a word made up of 3 tokens, it’s possible that each token is routed to a different expert. It’s just a model architecture.
But they are always training new models
Right, I mentioned that the more they’re used the more incentive there is to train new ones. But also it’s not like buying a product from the store which then creates a demand for a replacement, it’s much fuzzier and it’s difficult to point to any one user as increasing the demand for more models. Realistically I think we’ve gone well past the point where it makes sense to train general purpose LLMs any further, it’s drastically diminishing returns for marginally higher quality. The continued fervor is being driven by the same people that drove crypto and NFT prices way beyond reality: speculators and VC parasites trying to shove the new hype buzzword into anything they can get their hands on.
people who complain about mountains of disposable diapers at landfills misunderstand the problem. it’s the poop of newborn babies that’s the problem, not the ones who have gone on to become adults. although to be fair their growing up increases the chance of more poop-producing babies.
but yeah sleeping around without birth control isn’t contributing to even one extra poop-filled diaper.
Miyazaki’s sadness was enough for me. He is right. This is humans losing faith in humans. Trust the machine, not yourself.
Also AI is still worse than a human on things like essay writing. Why do I know? Cause I just finished grading midterms!
His popular AI quote is from 2016 and is missing a lot of context. What he was commenting on isn’t anything like the current generative AI wave. That being said, he doesn’t seem to have publicly rectified it so it might still represent his views.
deleted by creator
Don’t really see how that doesn’t relate. So its not a reinforcement learning model designed to make animations. Cool, the result is still the same. Humanity losing faith in itself quote really can’t be applied in a different way to only refer to this one specific model that was made to make terrifying animations, it clearly applies to handing all this human made work over to machines that dont understand why we make what we make. The machine, and subsequently the people who created it, were accused by Miyazaki of not knowing suffering. Not having any idea about something they were trying to emulate. This is what struck his core. The lack of empathy or connection to the subject. The root of all of our connections and bonds come from shared experience and empathy. He was speaking on the abandonment of these principles and AI is the epitome of it all.
Thank you, way too many people here who seem to completely misunderstand the nature of Miyazaki’s resentment towards AI.
He was not simply put off by the appearance of the animations, but rather repulsed by the entire process and the idea that machines could ever replicate the creativity of humanity. This is a man that had one of his animators work more than a year on a 4 second shot, refusing to use CGI in any capacity to speed that process up. The notion that he would have anything but contempt for AI is laughable.
That stuff Miyazaki said was before generative AI existed. He was commenting on procedural animation being used poorly in a 3D simulation. It’s fair to apply his sentiment to AI, but he himself was not talking about AI.
Those animations were cursed.
Oh yeah, the presentation he was commenting on did suck, and while what he said to those guys was harsh it was entirely justified.
Somebody said The Apple ads for AI look like they’re describing the people who are the biggest pieces of shit you work with or know.
There’s a misconception regarding the “consumption” of water, also a bit of a bias towards AI data centers whereas most used water is actually from energy production (via carbon, fuel or even hydroelectric) which is actually a factor to be considered when calculating the actual water use and consumption.
Regarding energy production and water “consumption” I read some papers and as far as I could understand numbers flactuate wildly. 5-40% of the water that runs through the system ends up being consumed via evaporation (so from potentially drinkable/usable for agriculture water to mostly water that ends up in the sea).
What I’m trying to say is that, yes, we should be very aware of the water that we consume in our big data centers but should also put a great focus on the water used by the energy that fuels the data center itself, much of the discourse ends up being “haha use water for email silly” when it should be a catalyst for a more informed approach to water consumption.
Basically I fear that the ai industry can make use of our ignorance and eappease with some “net zero” bs completely ignoring where most of the water is consumed and how.
And yes there are solutions to avoid using fresh water for energy production: solar/wind, using sea water, using polluted water, more sophisticated systems that actually “consume” as little water as possible. These methods have drawbacks that our governments and industry refuse to face and would rather consume and abuse our resources, I really want people to focus on that.
How does it actually consume the water?
Cooling of datacenters.
How does it consume water? I thought it would be a closed loop?
It doesn’t. “Wasting water” is bullshit most of the time. What you waste is the energy powering pumps and sewage plants.
“Wasting water” is bullshit most of the time.
Pumping water out of reserves, using it as coolant, and then disposing the hot water into local waterways where the heat kills off the local ecology is “waste” on several levels.
This is a common practice for industrial cooling, as pumping water and releasing it is cheaper than cycling the water through a large ventilator and recovering it.
Cooling a datacenter doesn’t “consume” water in any way.
http://arxiv.org/pdf/2304.03271
That’s not true.
What a mouthful of a PDF lol. But as far as I understand, that PDF has nothing to do with datacenter cooling. Cooling a datacenter usually happens in a closed loop, meaning there is no place the water could evaporate (which is the closest thing we have to “consuming water”) to, so there is no loss. The water is cooled via a heat exchanger, which is not opening up the loop. We have the same concept with AIOs on PCs, and you don’t have to refill the water every now and then, because it doesn’t evaporate.
The PDF refers to power production (as most sources of power do rely on water), where there is certainly some amount of loss. But that is not what I was arguing against.
Ren’s most recent work focuses precisely on how AI is increasing water use. A large language model like OpenAI’s popular ChatGPT-3 must first be trained, a data and energy intensive process that can also boost water use. Ren found that training GPT-3 in Microsoft’s high-end data centers can directly evaporate 700,000 liters, or about 185,000 gallons, of water.
Once the AI model is in use, each inference, or response to queries, also requires energy and cooling, and that, too, is thirsty work. Ren and his colleagues estimate that GPT-3 needs to “drink” a 16-ounce bottle of water for roughly every 10-50 responses it makes, and when the model is fielding billions of queries, that adds up.
The researchers are saying otherwise. I tend to believe them
It’s very cool that you tend to believe them, but I’d like to understand how something in a closed loop is “evaporating” - that is physically impossible. I once heard they are planning to build datacenters in the ocean, but even then evaporation is unlikely as the datacenters won’t boil the ocean. The only way to make this work is if they submerge it in a small pond/lake or just flood the building, and keep dumping water into it - which is stupid aswell because there are MUCH better materials for that that are NOT conductive, like special oils, which are not water based.
So ye, believing researchers is one thing, but believing something that physically is not possible because it fits your narrative is stupid.
Evaporation, is my understanding. Even sealed containers have evaporation in heat conditions.
I would be interested in seeing the power consumption required to generate for an AI vs an artist, on an individual basis it might not stack up the way people want.
Capitalist dystopia got us comparing ingested calories per unit of art
Fuuuck, comment of the day right here IMO. This hit me.
Maybe about 33% less electricity than human digital art? I don’t feel like calculating this myself.
https://www.reddit.com/r/aiwars/comments/11v5ovu/comment/jcsj7uy/
I have local AI for this reason. All it does is toast my balls a bit, and waste 10’s of watts of electricity.
I want cover letters to be shot in the street at noon
I mean it’s not about the convience of writing bullshit emails and generating fun pictures, that can be done locally easily, it’s about these “AI” companies being shit.
You’re not gonna save the world by not using ChatGPT, just like you won’t save all those slaves in Zambia by not buying from Apple, and just like you didn’t destroy Twitter by joining Bluesky.
To have real effect requires systemic change, so if you want to actually make a difference you can do things like canvassing, running for local office positions and school boards, educating friends and family about politics, or try killing a few politicians and tech CEOs. You know, basic stuff.
Also I asked Gemini’s Deep Research to research this for me because why not UwU
Executive Summary
Estimates for the energy consumed by ChatGPT during its training and inference phases vary considerably across different studies, reflecting the complexity of the models and the proprietary nature of the data. Training a model like GPT-3 is estimated to require around 1.3 GWh of electricity1, while more advanced models such as GPT-4 may consume significantly more, with estimates ranging from 1.75 GWh to over 62 GWh.2 Models comparable to GPT-4o are estimated to consume between 43.2 GWh and 54 GWh during training.3 These figures represent substantial energy demands, with the training of GPT-4 potentially exceeding the annual electricity consumption of very small nations multiple times over. The energy used during ChatGPT inference, the process of generating responses to user queries, also presents a wide range of estimates, from 0.3 watt-hours to 2.9 watt-hours per query.4 This translates to an estimated annual energy consumption for inference ranging from approximately 0.23 TWh to 1.06 TWh. This level of energy demand can be comparable to the entire annual electricity consumption of smaller countries like Barbados. The lack of official data from OpenAI and the diverse methodologies employed by researchers contribute to the variability in these estimates, highlighting the challenges in precisely quantifying the energy footprint of these advanced AI systems.4
You’re who the meme is about
You’re who my comment is about.
Also I demand that everyone who calls it AI instead of procedural generation gets tazed on the butthole
No, please no! “Computer generated” became a snarl word since genAI, and it’s still AI even if it’s not on human level.
It probably doesn’t consume as much. Cars however… 👀
Projected to consume 25% of total domestic energy production by the time the build out is done
Don’t cars account for 30% of total energy consumption, and 70% of petroleum? Not defending AI here, the last thing we need is another unchecked, massive, resource hog.
Don’t cars account for 30% of total energy consumption, and 70% of petroleum?
Transportation does. But that includes planes, trains, and boats, not just cars.
Not defending AI here, the last thing we need is another unchecked, massive, resource hog.
I’ve seen the argument that AI will reduce travel demand. But I’ve also seen AI guys lobby aggressively to end Work From Home.
where you get this number?
Arm CEO Rene Haas cautions that if AI continues to get more powerful without boosts in power efficiency, datacenters could consume extreme amounts of electricity.
Haas estimates that while US power consumption by AI datacenters sits at a modest four percent, he expects the industry to trend towards 20 to 25 percent usage of the US power grid by 2030, per a report from the Wall Street Journal.
How is it consuming so much water? It is probably used for cooling the data centers, but why could it not be reused or even used as heating network?
or they could run the models on their own machines and not cause environmental problems for the rest of us
