Having open GPU ISA's is not the same as requiring them all to be the same. GCC and LLVM both compile to a variety of architectures but manage to share a lot of

Having open GPU ISA's is not the same as requiring them all to be the same. GCC and LLVM both compile to a variety of architectures but manage to share a lot of code and optimizations between all of the targets. Having the same for GPUs, where the vendor gives you an ISA and you can use whatever compiler you like (even hand written assembly), would be a lot more preferable. This is what the open-source Gallium3D stack does, though it's hindered by relying on reverse-engineered information.