Ask Lemmy
A Fediverse community for open-ended, thought provoking questions
Rules: (interactive)
1) Be nice and; have fun
Doxxing, trolling, sealioning, racism, and toxicity are not welcomed in AskLemmy. Remember what your mother said: if you can't say something nice, don't say anything at all. In addition, the site-wide Lemmy.world terms of service also apply here. Please familiarize yourself with them
2) All posts must end with a '?'
This is sort of like Jeopardy. Please phrase all post titles in the form of a proper question ending with ?
3) No spam
Please do not flood the community with nonsense. Actual suspected spammers will be banned on site. No astroturfing.
4) NSFW is okay, within reason
Just remember to tag posts with either a content warning or a [NSFW] tag. Overtly sexual posts are not allowed, please direct them to either !asklemmyafterdark@lemmy.world or !asklemmynsfw@lemmynsfw.com.
NSFW comments should be restricted to posts tagged [NSFW].
5) This is not a support community.
It is not a place for 'how do I?', type questions.
If you have any questions regarding the site itself or would like to report a community, please direct them to Lemmy.world Support or email info@lemmy.world. For other questions check our partnered communities list, or use the search function.
6) No US Politics.
Please don't post about current US Politics. If you need to do this, try !politicaldiscussion@lemmy.world or !askusa@discuss.online
Reminder: The terms of service apply here too.
Partnered Communities:
Logo design credit goes to: tubbadu
view the rest of the comments
If you mean distributing inference across many machines, each of which could not individually deal with a large model, using today's models, not viable with reasonable performance. The problem is that you require a lot of bandwidth between layers; a lot of data moves. When you cluster current systems, you tend to use specialized, high-bandwidth links.
It might theoretically be possible to build models that are more-amenable to this sort of thing, that have small parts of a model run on nodes that have little data interchange between them. But until they're built, hard to say.
I'd also be a little leery of how energy-efficient such a thing is, especially if you want to use CPUs
which are probably more-amenable to be run in a shared fashion than GPUs. Just using CPU time "in the background" also probably won't work as well as with a system running other tasks, because the limiting factor isn't heavy crunching on a small amount of data
where a processor can make use of idle cores without much impact to other tasks
but bandwidth to the memory, which is gonna be a bottleneck for the whole system. Also, some fairly substantial memory demands, unless you can also get model size way down.