Energy Consumption for AI Images - How Much?

richardandtracy

There was an interesting radio programme yesterday on BBC Radio 4's 'Digital Human ' programme about the energy consumption by AI.

The main conclusion of the programme was that there was little information in the public domain, however what little was available appears somewhat concerning, indicating that if used widely it may lead to large power demand growth. Two examples could be found, as enumerated below:

1. A ChatGPT search for when a shop is open takes 'about 30x the energy of a standard Google search' for the same thing.
2. A 4k square rendered image takes 'The same energy as about half the charge of a mobile phone '

Note, in neither case was the programme able to find real world numbers for kW.hrs so it's really hopelessly vague and of no real practical use in determining behaviour.

On the other hand, DAZ must have an idea, as they're paying for the electricity bill on a server farm, and presumably the bill has changed since introducing DAZ AI Studio. Is there any real data DAZ could share with us to indicate whether the energy consumption reports of AI taking more power than conventional computing are based on fact or incorrect information?

That would be a really useful thing for DAZ to do. If AI image creation is not power hungry, there would be some nice data to come back to critics with. And if it is, maybe self rationing to 'important' images rather than 'at a whim' images could reduce the extra cost to DAZ.

Regards,

Richard.

WendyLuvsCatz

I use both Stable Diffusion in Automatic 1111 and DAZ studio on my PC

SD uses considerably less power and is much faster doing 1920x1080 images

richardandtracy

That's interesting. If anything, suggesting that it's an argument to migrate towards AI rather than being cautious - on the power consumption side, disregarding all else. I do wish there was sufficient information around to make informed choices. Regards, Richard.

WendyLuvsCatz

all information out there is biased and focusses on either the ethical issues with using the Laion5 database or the environmental effects of big data centres used for training and commercial uses, the latter exist anyway for Google, Microsoft, Apple, public and private enterprise and other internet services too be it for AI or other functions related to the internet even existing, just how much is actually dedicated to large language models and machine learning and how much more emission gasses would be produced doing the work it replaces manually is anyone's guess

wolf359

WendyLuvsCatz said:

all information out there is biased and focusses on either the ethical issues with using the Laion5 database or the environmental effects of big data centres used for training and commercial uses, the latter exist anyway for Google, Microsoft, Apple, public and private enterprise and other internet services too be it for AI or other functions related to the internet even existing, just how much is actually dedicated to large language models and machine learning and how much more emission gasses would be produced doing the work it replaces manually is anyone's guess

 

We stated Wendy.

FirstBastion

When I had a slower computer,  and renders would take ten minutes,  I often went brew a cup of coffee,  and thought about the 900 watt coffee maker using up the same amount of power as my 900 watt computer over that time  I had to pay the electricity on both.  As long as it is clean electricity,  be it renewable or nuclear,  faster computing will be a net positive gain.

generalgameplaying

richardandtracy said:

That's interesting. If anything, suggesting that it's an argument to migrate towards AI rather than being cautious - on the power consumption side, disregarding all else. I do wish there was sufficient information around to make informed choices. Regards, Richard.

If it's a can-do argument, ok. If it's about ecologic judgement, apparently the training is super heavy.

(Millions of users using the model then lead to an amortized thing per generation. With all the different models everywhere and constant retraining, it remains a big number, and it's still growing.)

huntermorrow3301

FirstBastion said:

When I had a slower computer,  and renders would take ten minutes,  I often went brew a cup of coffee,  and thought about the 900 watt coffee maker using up the same amount of power as my 900 watt computer over that time  I had to pay the electricity on both.  As long as it is clean electricity,  be it renewable or nuclear,  faster computing will be a net positive gain.

That’s the main idea! Whether AI or a massive futuristic city, energy use will keep growing, no matter what tech we use. Instead of only trying to cut down on energy (which might be impossible already), we should focus more on finding better energy sources.

Martirilla

Cost of weekly AI image generation in wH’s: https://sketchbooky.wordpress.com/2023/12/01/cost-of-weekly-ai-image-generation-in-whs/

richardandtracy

That, at last, looks like a plausible assessment of the real costs. I think we can say that the energy drain claim is de-bunked. It's more difficult to assess on a server farm, but there is little reason to believe it's orders of magnitude less efficient. Regards, Richard.

JD_Mortal

In the way of concept art, 2D art and photo-realism... This is my 2-cents.

It would take me a week to a month, to make similar art, by hand, on a computer. A task that takes a few minutes, with AI, to produce dozens of flavored results, well above my ability. The power consumption is minimal, for those tasks. Compounding the detail, with AI in-painting and up-scaling and deep-redetailing... Still just a few more minutes, what would take another month, by hand.

As for a 3D render, simulation result, styled to look like IRAY and DAZ3D, I am sure that it consumes less power than just rendering and setting-up a scene. Then re-rendering again, and again, and again, to get a better result. (For those of us a bit more challenged with posing and lighting and scene setup.)

Though, I can not, for the life of me, figure out why people want to use AI to generate "clearly fake people", within a "realistic scene". I would think that Daz would have spent more time trying to make the "clearly fake people", look like "real people", as a compliment for the shortfalls of daz, not an amplification of the biggest flaw in these 3D models. Most notable in the hair, faces, contour-wrapped clothing and unflexed posed bodies that all look like they should fall over and defy gravity, which clearly identifies even the best renderings as a "daz model".