Researchers Quantify the Carbon Footprint of Generating AI Images (engadget.com) 47
Researchers at the AI startup Hugging Face and Carnegie Mellon University discovered (PDF) that generating an image using artificial intelligence has a carbon footprint equivalent to charging a smartphone. Meanwhile, AI-generated text takes up as much energy as charging a smartphone to about 16 percent full. Engadget reports: The study didn't just look into image and text generation by machine learning programs. The researchers examined a total of 13 tasks, ranging from summarization to text classification, and measured the amount of carbon dioxide produced per every 1000 grams. For the sake of keeping the study fair and the datasets diverse, the researchers said they ran the experiments on 88 different models using 30 datasets. For each task, the researchers ran 1,000 prompts while gathering the "carbon code" to measure both the energy consumed and the carbon emitted during an exchange.
The findings highlight that the most energy-intensive tasks are those that ask an AI model to generate new content, whether it be text generation, summarization, image captioning, or image generation. Image generation ranked highest in the amount of emissions it produced and text classification was classified as the least energy-intensive task. The researchers urge machine learning scientists and practitioners to "practice transparency regarding the nature and impacts of their models, to enable better understanding of their environmental impacts."
The findings highlight that the most energy-intensive tasks are those that ask an AI model to generate new content, whether it be text generation, summarization, image captioning, or image generation. Image generation ranked highest in the amount of emissions it produced and text classification was classified as the least energy-intensive task. The researchers urge machine learning scientists and practitioners to "practice transparency regarding the nature and impacts of their models, to enable better understanding of their environmental impacts."
Watt seconds (Score:4, Informative)
Stable Diffusion uses a consumer GPU, and doesn't take all that long to generate an image. If it took 15 seconds, it would be consuming around 130W during that time. That doesn't quite sound like "charging a phone".
Re: (Score:3)
1 watt (power) is 1 joule (energy) per second (time). "Watts per second" would be how quickly power consumption increases or decreases. Energy consumption is 130 W for 15 seconds, or 1950 joules -- about half of a food calorie.
Re: (Score:2)
If people generate images instead of playing a video game, they reduce their CO2 emissions.
Re: (Score:2)
Yeah, because there's generally significant downtime between images. You do a batch, and then work on them in Gimp/Photoshop, then back to img2img,back to Gimp/Photoshop,etc. Maybe like 2-5% GPU usage for a particular workflow for me.
If one calls a typical GPU to consume 10x the power of a light bulb, then a 15s generation is the equivalent of having a single light bulb on for 2 1/2 minutes, or room in your house lit by five bulbs for 30 seconds-ish. That said, the GP's number of 15s seems abnormally lo
Re: (Score:3)
If we assume that a cell phone battery contains 10Wh, then 16% full would be 1,6 watt hours, which is a really minute amount of energy. 300W in 15 seconds (which again I contest is too much time) would be 1,25Wh. So the paper doesn't seem to be *radically* off - they might be assuming an unusually intensive generation and/or an unusually low-energy cell phone battery.
The problem is that I think most people grossly misunderstand how much energy it takes to charge a cell phone battery. 10Wh = said room in
Re: Watt seconds (Score:2)
To state this in the reverse, leaving room lights on burns more energy than you think.
I just wish someone could explain this to my teenager.
Re: (Score:2)
Just to clarify: my generation environment is a single 3090 power-limited to 300W (instead of the nominal 350W), because it saves you 17% on power but only costs 2-6% on performance.
A 4090 however should be significantly more efficient per unit compute than even my power-limited 3090. And modern server cards more efficient still.
Re: Watt seconds (Score:3)
Useless report.
Much more accurate to talk about energy consumption, not apply a generic conversion rate to go from a calculated energy usage that could be accurate to an unnecessary conversion that is based on politics.
Re: (Score:2)
They did say it was like charging a phone to 16%. That takes only a few seconds of charging, with most phones.
Re: (Score:2)
Re: (Score:2)
Your statement does not conflict with mine.
Re: (Score:2)
Re: (Score:2)
Your qualifier was "on newer phones." Sure, your Samsung Galaxy latest model or Titanium iPhone might take that long. But my Moto G (2021 model) does not take "hundreds of seconds" to get to that 16% mark, and it goes for about 3 days between charges. So no, my statement is not false just because your pricey latest-model phone takes longer to charge.
Re: (Score:2)
Everything yo
Re: (Score:2)
I have confidence, because I have the phone and have direct experience. Yes, it does take about two hours to charge, but that's because as it gets closer to full, the rate of charging slows down. That's how fast chargers work. You get a very fast rate of charge at the beginning, tapering off as the battery gets full. I find it annoying when people assume they know what they are talking about, basing their opinions on assumptions that they have not measured or researched.
Granted, if I'm using the GPS, it has
Re: (Score:2)
Again, this was such a strange hill for you to choose to die on but here we are.
Re: (Score:2)
You got a source for your data? Of course not, you're just making up your numbers from thin air. I've got actual timings.
Strange hill to die on? I don't see *you* backing off! So you're just as strange as me, apparently.
Re: (Score:2)
Re: (Score:2)
You really are full of yourself! You say it's easy, but you can't produce a single source yourself. Right.
Re: (Score:3)
You know, I think all these stupid pseudo-"units" serve primarily to make things sound larger than they are. Bad journalism, nothing else.
One sample (Score:2)
My office machine is an M1-based (gen 1) Mac Studio Ultra. Some numbers here:
I use DiffusionBee [diffusionbee.com] (a Stable Diffusion pre-trained ML variant.) Image generation of a quality result takes about 9 seconds with a 512x512 target; 5 seconds for a 256x256 target, both using the default model and default settings. So it looks linearly dependent on output
Should Be Marginal Carbon Footprint (Score:2)
Re: (Score:2)
That's a good point but if AI is going to put lots of people out of work we need to consider the carbon footprint saved to host and pay a human at work, and to transport him there, but then also include the welfare payments to have him unemployed and possibly the carbon cost of the maintaining the monetary and taxation system to make it possible.
This seems exhausting ... maybe we should get AI to do it.
Re:Should Be Marginal Carbon Footprint (Score:4, Funny)
As near as I can tell, if you want to be safe from any kind of AI based threat to your livelihood, you should become a hand model.
ROTFL (Score:2)
If only I had mod points, lol. [tips hat]
Re: (Score:2)
AI appears to reduce demand for knowledge workers, who might work from home, but increase demand for people who work on-premises. So the net effect might be more commuting rather than less.
Re: (Score:2)
The nature of these ML engines is you train them once, and then everything after that is near negligible in comparison; duplicating an already- trained instance requires nearly zero effort, for instance. So the training energy divides out into the number of prompt responses. It c
Re: (Score:2)
Yeah, last I saw, OpenAI was saying that only about 10% of their power was dedicated to training, 90% to inference.
Decelerationists consultants drolling at the $$ (Score:2)
But did they calculate... (Score:5, Interesting)
But did they calculate to carbon footprint of their research?
Did they include the carbon footprint of putting this clickbait on the internet and having people click on it??
What about the carbon footprint of the Slashdot discussion of the clickbait article about the research???
Re:But did they calculate... (Score:4)
I logged in for the first time this year to upvote this, but I have no karma. This is a fantastic comment - I am going to copy it into my notes so I can see in the future and laugh again and again.
Re: (Score:2)
Exactly. The vast majority of people on the planet, are motivated by their own ego and survival of their own family. That's not a bad thing, but it means 97% of all environmental and climate work and ideas are about self serving narratives. That means the narratives which win are the ones which bring more money and success to individuals, rather than the ideas which would actually work. That's just where we are as a species. We are ethically more evolved than what we were ten thousand years ago, but not tha
Re: (Score:2)
Sheer crazimess (Score:2)
Re: (Score:2)
So what we need is fewer breathers?
Re: (Score:2)
What's the carbon cost of jacking off to pr0n? (Score:2)
The footprint of taking a shit?
Clipping your toenails?
Changing your pronouns?
Whining about how some mundane activity is Killing The Planet?
Raising a middle finger?
Both?
Middle fingers *and* middle toes?
So what's the carbon footprint of using a human? (Score:1)
Because it's always the question, "compared to what?"
Humans eat, use power, drive cars... they use electricity on their laptop (80w) or computer (120w) that dwarfs that of a charging cellphone.
Of course I can generate images with solar power-- but no idea of what the footprint of my panels and 1.1kw battery is. Still it runs my laptop, lights, and a fan for 8 hours a day.
Re: (Score:2)
What they're saying is that ai generation isn't as free and over use poses costs.
Re: (Score:1)
Yes.. and what I'm saying is that human artwork generation maybe *much* more expensive than AI.
However, I agree with your point. Delivering those 6 images requires a lot of power. Each image would take *hours* to create on one PC with a single graphic card. Clearly any efficiency gains will have a large impact.
The actual conclusion... (Score:2)
They're basically saying that training each model to perform multiple types of tasks isn't particularly efficient. It sounds like what all these AI companies are showing off to the public at the moment are their all-singing, all-danci
Bullshit units (Score:2)
Bullshit summary, and bullshit article from Engadget. After reading the PDF from the summary I finally figured out that "measured the amount of carbon dioxide produced per every 1000 grams" was actually per 1000 queries.
Now that we have solved that mystery, what the hell does "charging a smartphone to 16%" even mean? This is Slashdot, we don't have to be scared of proper units like joules or watt-hours here. Again, looking further, the original authors of the original PDF also use these bullshit "smartphone
Re: (Score:2)
My office machine (which is the only one I run generative ML on) is 100% (locally, no utility company interaction) solar powered. So the carbon footprint is whatever the computer manufacture + ultracap + panel + charge controller carbon footprint was, divided by the number of ML generation instances. It's not zero. But it's also not increasing.