Technology

ChatGPT.

I’ve run an LLM on my desktop GPU and gotten decent results, albeit not nearly as good as what ChatGPT will get you.

Probably used less than 0.1Wh per response.

Yes, kind of… when those businesses making money out of the subscriptions are willing to ship with the OS for free which something only Apple has the luxury to do instead of OpenAI who doesn’t ship hardware or software (like Windows) beyond an app that’s less than 100MB. Servers would still be needed but not for general cases like help me solve this math or translation. Stable Diffusion or Flux is one example where you only need the connection to internet when downloading a certain model like you wouldn’t necessarily want to download every kind of game in the world when the intention is to play games arises.

Don't worry, we'll bail them out once the bubble bursts.

Billions of people have been using Google for years, and Google has been using artificial intelligence for years. It's nothing new.

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 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.

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?

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

Time to buy stock in water lol

Think of it like this: rich people accumulate more wealth by paying fewer people to accomplish more work faster, so it's worth burning through the worlds resources at breakneck speed to help the richies out, right?

It is a closed loop, but the paper treats it as if it's an open loop, and counts all water use for the building, as well as all the water that went into creating any equipment used.... and the water that escapes power plants in powering the buildings.... it also includes any other buildings that might house related services. Here is the original "study" which is about what maths could be done given the above assumptions:

http://arxiv.org/pdf/2304.03271

In short, it has nothing to do with reality, and is more just an attempt at the authors to get their names out there (on bad science that the media is interested in publicizing for click bait reasons).

Evaporative coolers save a ton of energy compared to refrigerator cycle closed loop systems. Like a swamp cooler, the hot liquid that comes from cooling the server is exposed to the atmosphere and enough evaporates off to cool the liquid by a decent percentage, then it's refrigerated before going back into the servers.

Data centre near me is using it and the fire service is used to be being called by people concerned the huge clouds of water vapor are smoke

Yes, the vast majority are closed loop systems and the water isn't really used up, like a lot of these headlines imply.

That's not to say the energy being used can't be put to better uses, though.