Erik Bernhardsson: evaluation

Transcript context

…Yeah, I think on the inference side, like that's where we like, you know, like from a cost perspective, like utilization perspective, we've seen, you know, like very good numbers and in particular, like it's our ability to start containers and stop containers very quickly. And that means that we can auto scale extremely fast and scale down very quickly, which means like we can always adjust the sort of capacity, the number of GPUs running to the exact traffic volume. And so in many cases, like that actually leads to a sort of interesting thing where like we obviously run our things on like the public cloud, like AWS GCP, we run on Oracle, but in many cases, like users who do inference on those platforms or those clouds, even though we charge a slightly higher price per GPU hour, a lot of users like moving their large scale inference use cases to model, they end up saving a lot of money because we only charge for like with the time the GPU is actually running. And that's a hard problem, right? Like, you know, if you have to constantly adjust the number of machines, if you have to start containers, stop containers, like that's a very hard problem. Starting containers quickly is a very difficult thing. I mentioned we had to build our own file system for this. We also, you know, built our own container scheduler for that. We've implemented recently CPU memory checkpointing so we can take running containers and snapshot the entire CPU, like including registers and everything, and restore it from that point, which means we can restore it from an initialized state. We're looking at GPU checkpointing next, it's like a very interesting thing. So I think with inference stuff, that's where serverless really shines because you can drive, you know, you can push the frontier of latency versus utilization quite substantially, you know, which either ends up being a latency advantage or a cost advantage or both, right? On training, it's probably arguably like less of an advantage doing serverless, frankly, because you know, you can just like spin up a bunch of machines and try to satisfy, like, you know, train as much as you can on each machine. For that area, like we've seen, like, you know, arguably like less usage, like for modal, but there are always like some interesting use case. Like we do have a couple of customers, like RAM, for instance, like they do fine tuning with modal and they basically like one of the patterns they have is like very bursty type fine tuning where they fine tune 100 models in parallel. And that's like a separate thing that modal does really well, right? Like you can, we can start up 100 containers very quickly, run a fine tuning training job on each one of them for that only runs for, I don't know, 10, 20 minutes. And then, you know, you can do hyper parameter tuning in that sense, like just pick the best model and things like that. So there are like interesting training. runs for, I don't know, 10, 20 minutes. And then, you know, you can do hyper parameter tuning in that sense, like just pick the best model and things like that. So there are like interesting training. I think when you get to like training, like very large foundational models, that's a use case we don't support super well, because that's very high IO, you know, you need to have like infinite band and all these things. And those are things we haven't supported yet and might take a while to get to that. So that's like probably like an area where like we're relatively weak in. Yeah. Have you cared at all about lower level model optimization? There's other cloud providers that do custom kernels to get better performance or are you just given that you're not just an AI compute company? Yeah.…

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