this post was submitted on 09 Jan 2025
483 points (99.2% liked)
Opensource
1708 readers
7 users here now
A community for discussion about open source software! Ask questions, share knowledge, share news, or post interesting stuff related to it!
⠀
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Running an llm llocally takes less power than playing a video game.
The training of the models themselves also takes a lot of power usage.
They are using open source models that have already been trained. So no extra energy is going into the models.
Of course, I mean the training of the original models that the function is dependent on. It's not caused by VLC itself of course.
Source?
I don't have a source for that, but the most that any locally-run program can cost in terms of power is basically the sum of a few things: maxed-out gpu usage, maxed-out cpu usage, maxed-out disk access. GPU is by far the most power-consuming of these things, and modern video games make essentially the most possible use of the GPU that they can get away with.
Running an LLM locally can at most max out usage of the GPU, putting it in the same ballpark as a video game. Typical usage of an LLM is to run it for a few seconds and then submit another query, so it's not running 100% of the time during typical usage, unlike a video game (where it remains open and active the whole time, GPU usage dips only when you're in a menu for instance.)
Data centers drain lots of power by running a very large number of machines at the same time.
Training the model yourself would take years on a single machine. If you factor that into your cost per query, it blows up.
The data centers are (currently) mainly used for training new models.
But if you divide the cost of training by the number of people using the model, it should be pretty low.
From what I know, local LLMs take minutes to process a single prompt, not seconds, but I guess that depends on the use case.
But also games, dunno about maxing GPU in most games. I maxed mine for crypto mining, and that was power hungry. So I would put LLMs closer to crypto than games.
Not to mention games will entertain you way more for the same time.
Obviously it depends on your GPU. A crypto mine, you'll leave it running 24/7. On a recent macbook, an LLM will run at several tokens per second, so yeah for long responses it could take more than a minute. But most people aren't going to be running such an LLM for hours on end. Even if they do -- big deal, it's a single GPU, that's negligible compared to running your dishwasher, using your oven, or heating your house.
Any paper about any neural network.
Using a model to get one output is just a series of multiplications (not even that, we use vector multiplication but yeah), it's less than or equal to rendering ONE frame in 4k games.
I know you are agreeing with me, but it being a "series of multiplications" is not terribly informative, that's basically a given. The question is how many flops, and how efficient are flops.
They aren't using a LLM, this is a speech to text model like Whisper.