The description of the CDNA 5 instruction set in TFA is interesting, because of the convergence with the NVIDIA ISA.

I mean, Nvidia says CUDA Tiles says CUDA Tiles MLIR is open source. https://news.ycombinator.com/item?id=46330732

Afaik, AMD, Google, and everyone else is still using some combo of mostly XLA, with StableHLO, IREE, etc? https://openxla.org/

I was going to lightly blow off more steam about AMD's TheRock all-in-one monorepo-ish thing for their graphics drivers being basically unbuildable on anything but a single distro, with little attention to tickets, but (in spite of this ticket being open) it looks like AMD did file and close quite a lot of sub tickets that resolve a huge amount of the issues. Very good to see this transition; I was really concerned that this was all a show release for very certain partners but not serious, but this looks like they are being pretty go-for-it. https://github.com/ROCm/TheRock/issues/3477

It is a little alarming to watch AMD try to move. On the one hand I think Senior Leadership did a very good very nice job setting forth some pretty solid hopeful goals. For day 0 Helios. They wrote it down, on github, which is such a hopeful good important differentiator, that I love. But, like, did their devs have good access to hardware, in quantity, to test and work on here? Why are these tickets so relatively quiet? Incredible ticket. Would that we could rely on anyone else in the world to do this. This is amazing. Of course, it's not done at all, and that looks bad, and I'm so scared to share it, in that light, but I just want to applaud so much, and suggest that this is how you break the moats. You open and share and talk. Also, like, get the work done. Get it working. But the range of tasks here is great too. AMD clearly sees what it takes. The IO acceleration stuff gives me confidence they understand the scope of the work, and that it's not just trying to slam a lot of gpu cores onto huge motherboards. https://github.com/ROCm/ROCm/issues/6335