Writing a 100-word email using ChatGPT (GPT-4, latest model) consumes 1 x 500ml bottle of water It uses 140Wh of energy, enough for 7 full charges of an iPhone Pro Max
Writing a 100-word email using ChatGPT (GPT-4, latest model) consumes 1 x 500ml bottle of water It uses 140Wh of energy, enough for 7 full charges of an iPhone Pro Max
140Wh seems off.
It’s possible to run an LLM on a moderately-powered gaming PC (even a Steam Deck).
Those consume power in the range of a few hundred watts and they can generate replies in a seconds, or maybe a minute or so. Power use throttles down when not actually working.
That means a home pc could generate dozens of email-sized texts an hour using a few hundred watt-hours.
I think that the article is missing some factor, such as how many parallel users the racks they’re discussing can support.
I like that the 140Wh is the part you decided to question, not the “consumes 1 x 500ml bottle of water”
An article that thinks cooling is “consuming” should probably be questioned in all its claims.
I think there’s probably something wrong with the math around per-response water consumption, but it is true that evaporative cooling consumes potable water, in that the water cannot be reused until it cycles through the atmosphere and is recaptured from precipitation, same way you consume water by drinking and pissing it out, or agriculture consumes it for growing things. Fresh water usage is a major concern and bottleneck, especially with climate change. With the average data centre using 300k gallons of water per day, and Google’s entire portfolio using 5bn gallons per day, it’s not nothing.
I would say a model like ChatGPT could use a bit more energy than 7B llama
The study that suggests 10-50 interactions with ChatGPT evaporates a whole bottle of water, doesn’t account for the fact that cooling systems are enclosed…
…and that “study” is based on a bunch of assumptions, which include evaporation from local power plants, as well as the entire buildings GPT’s servers are located in. It does this as if one user is served at a time, and the organizations involved (such as microsoft) do nothing BUT serve one use at a time. So the “study” (which isn’t peer reviewed and never got published) pretends those buildings don’t also serve bing, or windows, or all the other functions microsoft is involved with. It instead assumes whole buildings at microsoft are dedicated to serving just one user of ChatGPT at a time.
It also includes the manufacture of all the serve and graphics cards equipment, even though the former was used before ChatGPT, and will be used for other things as well… and the latter is only used in training.
You can check the study out yourself here:
http://arxiv.org/pdf/2304.03271
It’s completely junk. Worthless. Even uses a click bait title, and keeps talking about “the secret water foot print” as if it’s uncovering some conspiracy. It’s bunk science.
P.S It also doesn’t seem to understand that the bulk of GPT’s training was a one time cost, paid in 2021, with one smaller update in 2023.
You are conveniently ignoring model size here…
Which is a primary impact on power consumption.
And any other processing and augmentation being performed. System prompts and other things that are bloating the token size …etc never mind the fact that you’re getting a response almost immediately for something that an at home GPU cluster (not casual PC) would struggle with for many minutes, this isn’t always a linear scale for power consumption.
You are also ignoring the realities of a data center. Where the device power usage isn’t the only power consumption of the location, cooling must be taken into consideration as well. Redundant power switching also comes with a percentage loss in transmission efficiency which adds to power consumption and heat dispersion requirements.
It’s true, I don’t know how large the models are that are being accessed in data centers. Although if the article’s estimate is correct, it’s sad that such excessively-demanding models are always being used for use-cases that could often be handled with much lower power usage.
This seems a big waste of energy if that 140Wh (504,000 joules) number is correct. That amount of energy is about 2,000 times what it would take to do a very similar thing on a home PC.
Writing a 100 word email with a 7B model would take my PC about 5 seconds, times an increased power use of 50 watts, so 250 joules.
I get that they might be using a much larger model, but the e-mail is not going to be 2,000 times better.
Datacenter LLM tranches are 7-8 H100s per user at full load which is around 4 kW.
Multiply that by generation time and you get your energy used. Say it takes 62 seconds to write an essay (a highly conservative figure).
That’s 68.8 Wh, so you’re right.
Source: I’m an AI enthusiast
Does that account for cooling? Storage? Networking? Non-H100 compute and memory?
Nope. Just GPU board power draw. 60 seconds is also pretty long with how fast these enterprise cards are but I’m assuming they’re using a giant 450B or 1270B model.
kW is a unit of instantaneous power; kW/s makes no sense. Note how multiplying that by seconds would cancel time out and return you power again instead of energy. You got there in the end, though.
Woop, noted, thanks
Well that’s of the same order of magnitude as the quoted figure. I was suggesting that it sounded vastly larger than it should be.
They’re probably factoring in cooling costs and a bunch of other overhead, I dunno
That’s what I always thought when reading this and other articles about the estimated power consumption of GPT-4. Run a decent 7B LLM on consumer hardware like the steam deck and you got your e-mail in a minute with the fans barely spinning up.
Then I read that GPT-4 is supposedly a 1760B model. (https://en.m.wikipedia.org/wiki/GPT-4#Background) I don’t know how energy usage would scale with model size exactly, but I’d consider it plausible that we are talking orders of magnitude above the typical local LLM.
considering that the email by the local LLM will be good enough 99% of the time, GPT may just be horribly inefficient, in order to score higher in some synthetic benchmarks?
Computational demands scale aggressively with model size.
And if you want a response back in a reasonable amount of time you’re burning a ton of power to do so. These models are not fast at all.
Thanks for confirming my suspicion.
So, the whole debate about “environmental impact of AI” is not about generative AI as such at all. Really comes down to people using disproportionally large models for simple tasks that could be done just as well by smaller ones, run locally. Or worse yet, asking a behemoth model like GPT-4 about something that could and should have been a simple search engine query, which I (subjectively) feel has become a trend in everyday tech usage…