You're viewing a single thread.
Let’s do the math.
Let’s take an SDXl porn model, with no 4-step speed augmentations, no hand written quantization/optimization schemes like svdquant, or anything, just an early, raw inefficient implementation:
So 2.5 seconds on an A100 for a single image. Let’s batch it (because that’s what’s done in production), and run it on the now popular H100 instead, and very conservatively assume 1.5 seconds per single image (though it’s likely much faster).
That’s on a 700W SXM Nvidia H100. Usually in a server box with 7 others, so let’s say 1000W including its share of the CPU and everything else. Let’s say 1400W for networking, idle time, whatever else is going on.
That’s 2 kJ, or 0.6 watt hours.
…Or about the energy of browsing Lemmy for 30-60 seconds. And again, this is an high estimate, but also a fraction of a second of usage for a home AC system.
…So yeah, booby pictures take very little energy, and the usage is going down dramatically.
Training light, open models like Deepseek or Qwen or SDXL takes very little energy, as does running them. The GPU farms they use are tiny, and dwarfed by something like an aluminum plant.
What slurps energy is AI Bros like Musk or Altman trying to brute force their way to a decent model by scaling out instead of increasing efficiency, and mostly they’re blowing that out of proportion to try the hype the market and convince them AI will be expensive and grow infinitely (so people will give them money).
That isn’t going to work very long. Small on-device models are going to be too cheap to compete.
https://escholarship.org/uc/item/2kc978dg
So this is shit, they should be turning off AI farms too, but your porn images are a drop in the bucket compared to AC costs.
TL;DR: There are a bazillion things to flame AI Bros about, but inference for small models (like porn models) is objectively not one of them.
The problem is billionaires.
I don’t disagree with you but most of the energy that people complain about AI using is used to train the models, not use them. Once they are trained it is fast to get what you need out of it, but making the next version takes a long time.
Only because of brute force over efficient approaches.
Again, look up Deepseek's FP8/multi GPU training paper, and some of the code they published. They used a microscopic fraction of what OpenAI or X AI are using.
And models like SDXL or Flux are not that expensive to train.
It doesn’t have to be this way, but they can get away with it because being rich covers up internal dysfunction/isolation/whatever. Chinese trainers, and other GPU constrained ones, are forced to be thrifty.
And I guess they need it to be inefficient and expensive, so that it remains exclusive to them. That's why they were throwing a tantrum at Deepseek, because they proved it doesn't have to be.
Bingo.
Altman et al want to kill open source AI for a monopoly.
This is what the entire AI research space already knew even before deepseek hit, and why they (largely) think so little of Sam Altman.
The real battle in the space is not AI vs no AI, but exclusive use by AI Bros vs. open models that bankrupt them. Which is what I keep trying to tell /c/fuck_ai, as the "no AI" stance plays right into the AI Bro's hands.
people complain about AI using is used to train the models, not use them
that's absolutely not true. In fact, most people who complain don't even know the difference.
Thx for doing the math
Not only are they cheaper than AC, but doing the math shows that they are more energy efficient than a human doing the same work, since humans operate at around 80-100W, 24 hours a day. (Assuming that the output is worth anything, of course.)
let's not use the term "efficiency" with humans making art, please. you're not helping anyone with that argument, you're just annoying both sides.
Well if humans could run on coal it would be a valid argument...
Humans essentially do run on fossil fuels. Modern agriculture is very energy intensive.
Humans at least run on renewable energy.
The computer you draw your art on, not so much. Reject modern art, embrace traditional carvings and cave paintings!
I think that’s going a bit far. ML models are tools to augment people, mostly.
Oh for sure. But if (for example) an artist can save time by tracing over an SDXL reference image, that is energy-efficient as well as time-efficient, despite most people claiming the contrary.