ROCM is a technology that is still in its early stages on Windows. Just a few days ago, a small portion was released, but it certainly requires a lot of work and development, which will most likely take the company more than a year to turn it into something decent and usable.
If we compare it to CUDA, even on Linux, there are numerous issues. Graphics cards in the field of artificial intelligence are lacking and leave much to be desired. If you're only considering changing your card due to VRAM, let me tell you that, for instance, in AI work environments like stable diffusion, an N.VID1A graphics card with half the RAM of an AMD one performs much better and has better workload management than one with double the RAM from AMD.
Even in Linux, where ROCm has been in development for many years, there are many instability issues that make solving certain problems with specific technologies very challenging.
I have an RX 6750 XT with 12GB of VRAM, and I've encountered too many issues with stable diffusion. Not to mention Windows, where practically nothing can be done. And the little that can be done is absurdly slow – it's like 1% of what you could achieve with Linux. And even when they released ROCm for Windows, there's no compatibility for many technologies – take a look at the documentation.
It depends on your work environment, but if I were you, I'd get a 4070 Ti or a 4070. Just the fact that they represent a generational leap in hardware for AI purposes makes them infinitely superior to our cards.
In my specific case, my card competes with a 3050 Ti – quite unfortunate, isn't it? After my experience with various software, I feel that AMD might be fine when you have very little money and only care about gaming (without raytracing).
