AMD already has major ongoing projects with OpenXLA/IREE. Lots of established engineers/researchers, and it’s in collaboration with Google/AWS. Hotz is delusional if he thinks that he can do better by ripping off Karpathy’s toy autograd implementation.
> AMD already has major ongoing projects with OpenXLA/IREE.
And how's that been going? The AMD stock price compared to NVidia seems to speak volumes about the efficacy of these projects.
IREE has been around for 5 years, without producing anything overtly practical. They seem to be focused more on academic jobs and citations. It's also focused on the general case of a compiler for "all" AI-type tasks, supporting everything from WASM to CUDA.
OpenXLA seems to be a bit more practical, but I spent the last 2 hours trying to make it work on my AMD card (Radeon Pro W7900) and failing.
I personally don't like Tinygrad's approach of doing their own thing rather than integrating into PyTorch/JAX/..., but it at least is _practical_ with a reasonable end-goal. Is it going to be successful? Who knows. But it's more practical than anything AMD has done within the recent 5 years.
I am an ML scientist, my company and several others are using IREE to deploy our models to edge devices. It is the most promising technology in this area.
Those academic publications are a sign that the people involved actually know what they’re doing, and are making sure their work holds up to scrutiny.
How does Tinygrad fall short? Performance is fine [0]. It's much smaller than Pytorch and all, but that's kind of in the name.
I've been hearing about MLIR and OpenXLA for years through Tensorflow, but I've never seen an actual application using them. What out there makes use of them? I'd originally hoped it'd allow Tensorflow to support alternate backends, but that doesn't seem to be the case.
