Creating a single AI-generated image needs as much power as charging your smartphone

Using a text-to-image model to craft an AI-generated image can require almost the same amount of power as that required to charge a smartphone, according to recent research. In a paper released on arXiv last week, a team of researchers from Hugging Face and Carnegie Mellon University calculated the amount of power AI systems …

[HuBo]

Do charge a thousand phones instead

Human artists are so much more creative and efficient than the giant energy sucking sound made by gen-AI ... why even bother!

[MacroRodent]

Re: Do charge a thousand phones instead

The human artists can be assumed to live and breathe anyway, whereas the generated AI images are an extra load.

[vtcodger]

Re: Do charge a thousand phones instead

"Human artists are so much more creative and efficient than the giant energy sucking sound made by gen-AI ... why even bother![?]"

Because humans are not the NEXT BIG THING. We humans may be energy efficient, but we are yesterday's news. This is not an issue of efficiency, it seems to be an issue of venture capitalists and other parasitic lifeforms seeking big bucks without having to work too hard at acquiring them.

[teknopaul]

Re: Do charge a thousand phones instead

I call BS on (11.49kWh) per image. Stable diffusion is giving me £1 per image as a free giveaway and all these AI image sites paid for by ads are getting £1 per click.

Something in these numbers don't add up. Maybe it's my price per kwh

[druck]

Re: Do charge a thousand phones instead

If you read the article, it's for a run of 1000 images.

[Anonymous Coward]

No free lunch

How does generation cost compare to the amortized cost of training?

[jeffdyer]

Power or energy? Disappointing reporting

[jdiebdhidbsusbvwbsidnsoskebid]

Power or energy? Disappointing reporting

Yeah, the general thrust of the article is fine but it looks like it's an early draft that hasn't been through review.

Change "power" for "energy" in the headline and probably everywhere else in the article as well. "Power" is fine when taking qualitatively about energy use generally, but not when you are getting specific and actually citing energy figures. To be fair on The Reg, the original paper has "power" and "watts" in the title, uses phrases like "power hungry" and then presents data in energy units.

The graph is described as measured CO2 emissions, but it's not measured, only modelled. The labelling on the graph even says so. Again, the paper includes the words "measure ... the carbon emitted"

The word "image" (I presume) is missing from the first sentence of paragraph 4.

I'm being harsh, the original paper is quite sloppy and it's maybe unfair to expect The Reg to do the critical paper review rather than just report the headline.

[Johannesburgel12]

They ran all tests on a single A100-80 on AWS. Those chips can go up to at most 500 Watts, assuming AWS stil runs the air-cooled variant (hardly any cloud provider has liquid cooling) it's more like 400 Watts peak. Even factoring in all the other components in the server, the PuE etc. it takes at least 4 hours to consume 3 kWh.

Aren't these image generation models much faster?

Also the A100-80 is not the king of inference.

[Pigeon Post]

£2.58 ?

Must be getting old. I read your sub-headline as "Microsoft to invest £2.58 in UK datacenters to power AI". Yay! I thought, common sense prevails...

But no...

[NeilPost]

Re: £2.58 ?

Wondering if GBP£2.5Bn and 20,000 GPU’s will be enough to A.I. compute power find any Brexit benefits ??

[ewan 3]

Nonsense

Given I can run stable diffusion XL on my home pc which is about 500-600w flat out and generate a picture in less than a minute, this seems inaccurate (0.008KWh). My phone takes at least an hour on a 30w charger (0.03KWh).

I get they say the most inefficient model, but that's stupid - use a representative one.

[Maryland, USA]

If you're trying to guilt me out of looking for AI porn, you'll have to do better.

[Omnipresent]

How much for a whole audio track?

Google just released AI music. Welcome to hell EDM kids. Everything you have ever done is irrelevant. https://artsandculture.google.com/experiment/8QFo2oQr2uT3pg

[Tom Womack]

Running DiffusionBee on my M1 Mac mini while looking at PowerMetric in another window, I'm getting about 9.5 watts of CPU+GPU power usage for about thirty seconds per image generated, so 0.00008 kilowatt-hours.

[Reginald O.]

Another new low...

"Sports Illustrated, came under fire last week for publishing AI-generated stories under fake bylines that included AI-generated photos and made-up bios for journalists that don't exist."

Bullshit sinks to a new low. Of course, 'they' will all be doing it soon. And, charging top dollar for the honor of being deceived. Who can you trust anymore?