AI image uses as much energy as a phone charge — are ChatGPT & co. harmful to the climate?

[Redaktion]

AI models like ChatGPT and Stable Diffusion require large amounts of energy and CO₂. This is a worrying trend, and recent studies provide important insights into the climate impact of artificial intelligence.


https://www.notebookcheck.net/AI-image-uses-as-much-energy-as-a-phone-charge-are-ChatGPT-co-harmful-to-the-climate.785320.0.html

[not_anton]

I create AI images on my iPhone/Mac with "Draw Things" app. It may use 3-5% of the charge per image but nothing like a full charge. Optimized data center GPUs that create images in large batches are even more efficient.

[vertigo]

AI will likely also result in lower power consumption by making tasks more efficient and helping advance technology faster. It will also result in advances in medicine and safety that will make people live longer, therefore making each person have a larger carbon footprint, which will almost certainly have much more impact than this. There are tons of factors weighing the scales in either direction. It could also be argued that computing in general creates a lot more carbon emissions than not having them. This is really no different. I keep seeing articles like this, yet there appears to be no concern about the use of induction charging, which is inherently less efficient, for Teslas, the whole point of which was supposed to be to reduce carbon emissions, yet this new charging method to avoid people having to go through the hassle of plugging/unplugging their car is going to waste electricity and partially negate that. Bottom line is things are almost never simple and almost always have numerous considerations and pros and cons.

[NikoB]

It will also result in advances in medicine and safety that will make people live longer

For the rich, yes, for everyone else - death exactly upon retirement.

Developed expert systems based on neural networks will be primarily used by those in power to strengthen effective control over the bulk of the population, increase their influence on the crowd, for the sake of their selfish interests and prolong their power indefinitely through their heirs. And not at all to improve the living conditions of the majority of the population. That excess part of the population that does not meet their direct interests will, on the contrary, be gradually and forcibly reduced by various methods.

Only naive children, with stereotyped thinking deliberately instilled in public schools, can still believe in a "bright future" for the bulk of the population...

[A]

I create AI images on my iPhone/Mac with "Draw Things" app. It may use 3-5% of the charge per image but nothing like a full charge. Optimized data center GPUs that create images in large batches are even more efficient.

This

[hugh]

Mmkay so then the 400 watt nvidia gpus and laptops with 150w gpus are "efficient" then? Do you think that electric cars are efficient where they have less utilty but require battery materials you throw away and are just as or more intensive than your 150 year old ICE that could do 50+mpg easily if further developed? Solar farms are hugely expensive and wasteful. And no saying some long term cataclysm is going to take place due to about 100 years or LESS of data is insanity, don't parrot mindlessly this PTB mantra about climate change, cc, 33. IF there are disasters this is well known climate ENGINEERING going on. Just cloud seeding is well known, and the public stuff is the TIP of the iceberg, sheesh dumbos.

[Neenyah]

Mmkay so then the 400 watt nvidia gpus and laptops with 150w gpus are "efficient" then?

Do people realise that stated TDP is not something that's always being used at that exact number? My GPU (6800 XT) has a TDP of 300 W and I barely even go above 130 W in most situations. Just like a 800 hp car isn't using 800 hp from the moment you turn its engine on.

[A]

Mmkay so then the 400 watt nvidia gpus and laptops with 150w gpus are "efficient" then?

You have to convert to Watt-hour first for finite-time tasks, because it's not running forever, 450W 4090 is generating 75 512x512 images per minute and 30 768x768 images per minute, so (if it's working at 100% power load) actual energy used is 0.1Wh and 0.25Wh per generation respectively - and these are the numbers you can compare across devices and battery capacities.
In the long run these numbers will be very efficient compared to other hardware, if not the most efficient (i think specialized TPUs/NPUs will beat it, back in 2021 Google TPUs were constantly beating A100 GPU I've used in efficiency).

[A]

P.S. Oh, and 4090 does 13 1024x1024 images per minute, that's 0.58Wh per generation.

[RobertJasiek]

For that purpose, account a whole computer - not just the GPU.

[A]

For that purpose, account a whole computer - not just the GPU.

The rest of the system is idling, even if you add 50W almost nothing will change, it'll just go from 0.1Wh to 0.11Wh