[mnau]

HIP was such a self-own and clear demonstration of AMD software capabilities... well, the lack of software capabilities. HIP was hard-coded for one GPU architecture. CUDA did it right, it has a intermediate virtual assembly PTX and driver compiles it to whatever actual instruction set card actually uses.

Imagine a meeting where they signed off on that. So each developer will have to provide a different binary for each of our architectures? Yep. And once we release the new architecture, developer will have to recompile his program for the new architecture? Yep. Sounds good to me.

[hgoel]

Yes, that was in part why they've had such a terrible history with GPU support.

They lost me as a customer when they rushed dropping support for the Radeon VII because of the need to ship binaries for every ISA, and didn't deliver proper 5700XT support until it was outdated.