My impression is that the accelerated computing side of AMD is receiving far, far too little attention. For example, their flagship GPU is still not officially supported by ROCm (AMD's answer to CUDA) [1]. Imagine the 2080 Ti not being supported by CUDA.
I've become a huge AMD fan, both because of their hardware, and because of their commitment to open source. But while the battles they have one against Intel on the x86 side are impressive, it seems that CUDA is leaving them far behind.
,,AMD ROCm is validated for GPU compute hardware such as AMD Radeon Instinct GPUs. Other AMD cards may work, however, they are not officially supported at this time.''
It seems like Lisa Su thinks that there's a separate ,,gamer market'' and ,,accelerator market''.
Jensen Huang understands that the same person can like to play games and train machine learning models on the same machine.
I'd love to switch to AMD CPU to have a portable laptop with low resource usage, as I spend most of my time travelling, but as GPUs in the cloud are overpriced (thanks to Jensen with separated pricing for servers), and internet in hotels are unpredictable, I don't want to train models in the cloud.
Anyways, Lisa said that she reads all comments about AMD, so I hope she'll listen :)
It likely will never have support. AMD chose to bifurcate their GPU designs into compute (CDNA) and games (RDNA) lines, with different architectures. RDNA sheds all the fancy features needed to support modern compute, thus gets more efficient in games that do not use it, but also cannot support the modern compute APIs.
NVIDIA is adding more features, like super-scaling to games, and machine learning models are improving faster than Moore's law. I expect those fancy features, like tensor cores to be a must for 4K gaming in the future.
What's funny is that the same strategy (leaving out specialized instructions from consumer level hardware) that worked extremely well for CPUs won't work for GPUs in my opinion.
If you look at ray tracing hardware (I have it on my RTX 2070 Max-Q card in my laptop), it sucks right now, but it's improving very fast as machine learning algorithms improve.
One thing that I forgot is that AMD can just focus on inferencing hardware (INT16 operations), and leave out tensor cores...so actually you are right, I'll just stay with NVIDIA GPUs.
I've become a huge AMD fan, both because of their hardware, and because of their commitment to open source. But while the battles they have one against Intel on the x86 side are impressive, it seems that CUDA is leaving them far behind.
[1] https://github.com/RadeonOpenCompute/ROCm/issues/887