CUDA, every mention
11 scenes · ← back to CUDA
tap a year for its mentions
every year anyone Stuart (Stu) 2Nick Joseph 1Greg Brockman 1Francois Chaubard 1Ankit Gupta 1
Verbatim, from the transcripts: the passages where CUDA comes up
Multi-GPU Kernels, Intelligence per Watt, Heterogeneous Inference, and More | YC Paper Club · Y Combinator
- ▶ 1:06 Francois Chaubard You can't do that with inference, and there's so many things like that that we'll talk about, and there's so much juice left to squeeze on the CUDA side, on the kernel side,
- ▶ 7:30 Stuart (Stu) Um, one quick note before I begin, um, the final deliverable of this paper is a CUDA framework that we refer to as parallel kittens, but the goal of today's talk is not to promote my open source library, but rather to convey the set of…
- ▶ 19:14 Stuart (Stu) So Parallel Kittens builds on all the trade-offs and principles that we discussed, and it's a highly opinionated set of CUDA programming primitives that extends Thunder Kittens, which is one of our previous works for single GPU kernels.
- ▶ 33:03 unnamed speaker Ultimately, these dispatch down to individual CUDA kernels that may or may not be very fast on modern hardware, which is why we get programming languages like Triton, which are a tile-based. 2 times in the scene
- ▶ 34:44 unnamed speaker And at least, like, for this archetype of people that just want state-of-the-art perf, as far as I can tell, like, they really prefer, like, CUDA, and it was only when we were doing lots of gem-related problems that people loved using…
- ▶ 35:32 unnamed speaker I've never written an open and CUDA book before.
- ▶ 37:23 unnamed speaker I'm gonna make you, if you've never written a CUDA kernel before, you're gonna do one with me right now.
- ▶ 1:14:36 unnamed speaker Like, I see a lot of programming language systems are kind of trending towards, okay, it's still the Cuda programming model, but it's Python syntax on top, which is great, like, simplifies things, easier to look at, but it doesn't actually…
Anthropic Head of Pretraining on Scaling Laws, Compute, and the Future of AI · Y Combinator
- ▶ 14:47 Ankit Gupta But not necessarily at the level of abstraction of, you know, writing custom CUDA kernels, or like was that also in the space where you guys were thinking about things?
- ▶ 51:16 Nick Joseph I was, like, working at the torch.matmol, but, like, I didn't know CUDA.
Building Dota Bots That Beat Pros - OpenAI's Greg Brockman, Szymon Sidor, and Sam Altman · Y Combinator
- ▶ 4:03 Greg Brockman And so what we've seen is that, well, you move your neural networks from running on a CPU to a GPU, and now suddenly you have a thousand CUDA cores running in parallel, uh, and that you can get massive performance boosts there.