> there is also second mover advantage, where they just need to copy-cat the best part of nvidia's ML stack and dont waste time figuring out what works/what doesn't work.
AMD doesnt have to implement CUDA api, they just need to make sure their compute framework works well with pytorch/tf/MLIR or whatever high level framework is being used.
Cuda itself will change over time, so no reason for AMD to pick cuda, because nobody writes CUDA kernels by hand, they use high level frameworks
Seems like patents would stop that.