Earlier quoted context omitted.
A better solution would be for AMD to invest in bringing their ML stacks up to date to work with PyTorch and such.
That would be nice. It really boggles the mind how thoroughly AMD has missed the boat on ML. And it really seems like they aren't on any trajectory to catch up even today. I've given up on them. Instead I'm hoping that Intel's imminent entry into the discrete GPU market does better. Nvidia is in desperate need of some competent competition in ML.
This was back in AMD's Bulldozer days (2011), when the company struggled both financially and technologically.
Meanwhile NVIDIA sponsored universities with graphics cards and had already developed their CUDA ecosystem (in 2007) when AMD was still busy with the ATI acquisition. In 2011 NVIDIA had an annual net income of about $500 million, while AMD had a net loss of $600 million at the same time and kept struggling for the following 5 years.
In other words, NVIDIA already had an existing ecosystem of professional grade H/W accelerators and S/W infrastructure, when AMD was still a CPU manufacturer without a dedicated GPU division. When AMD acquired ATI, NVIDIA was already in the process of transitioning their GPGPU stack from data centre-only products to consumer hardware.
AMD has powerful ML hardware today, but that's data centre and supercomputer only. They didn't miss the boat on ML - they were busy not drowning while NVIDIA was handing out goodies to academia.