Abhinav Venigalla: prediction

Transcript context

…Amazing answer. Well, thanks. Uh, we can, we can touch on a little bit [00:48:00] on, uh, efficiency and speed because we, we, uh, didn't mention about that. So right now people spend between three to 10 days. You, you spend 10 days on, on mpc, seven rep spend three days. What's feasible? What's what Do you wanna get it down to? Oh, for, for these original models? Yeah. Yeah. So I think, um, this is probably one of the most exciting years, I think for training efficiency, just generally speaking, because we have the, the combination of a couple things, like one is like this next generation of hardware, like the H 100 s coming out from Nvidia, which on their own should be like, at least like a two x improvement or they 100 s on top of that, there's also a new floating point format f P eight, um, which could also deliver that alone. Does it? Yes. Yeah. Yeah. How, what, why? Oh, the f p thing? Yeah. Yeah. So basically what's happening is that, you know, when we do all of our math, like in the models matrix, multiplication, math, we do it in a particular precision. We started off in 32 bit precision a few years ago, and then in video came with 16 bit, and over the course of several years, we've all figured out how to do 16 bit training and that basically, you know, due to the harder requirements like. Increase the throughput by two x, reduce the cost by two x. That's about to happen again with FBA eight, like starting this year. And with Mosaic, you know, we've already started profiling L L M training with f p eight on H 100 s. We're seeing really, really good improvements there. And so you're gonna see a huge cost reduction this year just from this hardware fact alone. On top of that, you know, there's a lot of architectural applications. We're looking at ways to introduce some forms of sparsity, not necessarily like the, the, the super unstructured sparsity like lottery ticket. Um, which not that I'm sure I'm really happy to talk about. Um, but, but, um, are there ways of doing, like you gating or like, kind of like m moe style architectures? So, you know, I think originally, you know, what was like 500 k. I think to try and train a Jeep, the equality model, if at the end of the year we could get that down to a hundred k, that would be fantastic. That is this year's type of thing.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.