Samsung's fabs have been getting better and better over the years, we could see a surprise blowout on the 3nm node. Hell, just the other day Samsung was testing an RDNA2 iGPU that was faster than the GPU on the iPhone 12... on the 7nm node. I'm no expert here, but AMD might finally have the silicon it needs to compete with Nvidia in raw compute.

That’s exciting news, but I can’t help but feel that at least in my field (machine learning), Nvidia is far more sticky than just their compute dominance. As long as CUDA is Nvidia proprietary, I just don’t think our team can afford to move.

I’m seeing lots of very impressive movements from Tensorflow and Pytorch to support ROCm, but it’s just not at a level yet where it would make good business sense for us to switch, even if AMD GPUs were 50% faster than Nvidia. And it seems like Nvidia is improving and widening their compute stack faster than AMD is catching up.

I'm right there with you, I have an Nvidia GPU on all my machines too. With that being said, however, there are plenty of software-agnostic workloads that I run that could benefit from a truly open, powerful GPU.