It's not just CUDA vs ROCm, ROCm has come a long way and is pretty compelling right now. What they lack is the proper hardware acceleration (e.g tensor cores).
And getting ROCm set up is still a buggy experience with tons of fiddling in the deep inner workings of Linux so it is nearly impossible for the average machine learning engineer to use.
It is compelling for certain bespoke projects like Europe's shiny new supercomputer, but for the vast majority of machine learning, it is totally unusable. By now in ML world the word "gpu" is synonymous with "nvidia".
We actually had more issues with nvidia drivers messing up newcomers' machines during updates than with setting up AMD GPUs, but then again n is small (and AMD GPUs were for playing around rather than real work).
Still, a Titan Xp has CUDA support and plenty of memory, but it's better, IME, to upgrade to a model with less memory but higher cuda compute and access to tensor cores.
For the amount invested in the hardware development, the amount AMD have been investing in the software side has been shocking.