> Nvidia is the only one that wants vendor-specific code.

Isn't that because CUDA is better and that tight software/hardware integration is more powerful?

If it wasn't then presumably people would be using AMD GPUs for their deep learning, but they're not.