Is there anyway to make it use less at it gets more advanced or will there be huge power plants just dedicated to AI all over the world soon?
Is there anyway to make it use less at it gets more advanced or will there be huge power plants just dedicated to AI all over the world soon?
imagine that to type one letter, you need to manually read all unicode code points several thousand times. When you’re done, you select one letter to type.
Then you start rereading all unicode code points again for thousands of times again, for the next letter.
That’s how llms work. When they say 175 billion parameters, it means at least that many calculations per token it generates
I don’t get it, how is it possible that so many people all over the world use this concurrently, doing all kinds of lengthy chats, problem solving, codegeneration, image generation and so on?
that’s why they need huge datacenters and thousands of GPUs. And, pretty soon, dedicated power plants. It is insane just how wasteful this all is.
So do they load all those matrices (totalling to 175b params in this case) to available GPUs for every token of every user?
yep. you could of course swap weights in and out, but that would slow things down to a crawl. So they get lots of vram (edit: for example, an H100 has 80gb of vram)
I also asked ChatGPT itself, and it listed a number of approaches, and one that sounded good to me is to pin layers to GPUs, for example we have 500 GPUs: cards 1-100 have permanently loaded layers 1-30 of AI, cards 101-200 have permanently loaded layers 31-60 and so on, this way no need to frequently load huge matrices itself as they stay in GPUs permanently, just basically pipeline user prompt through appropriate sequence of GPUs.
I can confirm as a human with domain knowledge that this is indeed a commonly used approach when a model doesn’t fit into a single GPU.