|> Well IMHO, the repeating theme was Nvidia is focused on one strategy and one strategy only, CUDA based GPU design|
But, if machine learning (and GPGPU, in general) really takes off, does the industry really want a proprietary solution? As far as I understand, a big part of the CUDA lure is not the underlying architecture, nor the programming language adaptations, but rather the libraries that lies on top. I think the industry will create competing libraries and adapt existing frameworks as we see alternative architectures for machine learning emerge.
AMD has done a lot already to establish an alternative open ecosystem, with HSA and ROCm. Hopefully, they will keep pushing the software forward and get some traction within the industry.