[xnx]

Is it possible someone will write a CUDA to AMD/Tensor/whatever transpiler (high-level emulator? I'm not sure of the right term) I thought there were a remarkably small number of ops that GPUs perform. Seems like a very high premium to pay for not wanting to rewrite in JAX or whatever.

[smallmancontrov]

Lack of translation layer wasn't the problem, AMD's shit being completely broken (for this purpose) was the problem. Driver-level broken, possibly hardware-level broken. Black screens, restarts, wrong answers, threads full of people shouting into the void year after year about the same behavior before seeing the light and buying nvidia.

Then NVDA pumped by trillions and AMD did not and even AMD's crack team of trained denial specialists could no longer stay the course, so they started to turn the ship. But that was only a year or two ago and even the tiniest changes take years in hardware land so the biggest sins are still baked in to the latest chips, but at least the software has started to suck a bit less.

I no longer see 100% of the people who try to build on AMD compute run away screaming. Many do, but not all. That's a change, an important and positive one. If AMD keeps it up maybe they can save us from 80% margins on matrix multiplication after all.

[drexlspivey]

AMD hired 1 guy to do it and then fired him

[xnx]

I wondered if John Carmack would take his strong knowledge of low level hardware and interest in AI and work on this, but he seems to be working on a non LLM flavor of AI.

[bayindirh]

From his talks, he never liked AMD/ATI hardware. Doom 3 code was infamously NVIDIA optimized, since he liked the hardware more.

[newsclues]

Or Abrash?