I also think something important here is AMD's strategy with APU has been small to large. Something that really stood out to me over the last few years is that NVidia was capturing the AI market with big and powerful GPU while AMD's efforts were all going into APU research at the low end. My belief is that they were preparing for a mobile-heavy future where small, capable all-purpose chips would have a big edge.
They might even be right. One of the potential advantages of the APU approach is if they GPU can be absorbed into the CPU with shared memory, a lot of the memory management of CUDA would be obsoleted and it becomes not that interesting any more. AMD are competent, they just have sucky crash-prone GPU drivers.
When I run a LLM (llama.cpp ROCm) or stable diffusion models (Automatic1111 ROCm) on my 7900XTX under Linux, and it runs out of VRAM, it messes up the driver or hardware so badly that without a reboot all subsequent runs fail.
You're probably using it for graphics though; the graphics drivers are great. I refuse to buy a Nvidia card just because I don't want to put up with closed source drivers.
The issue is when using ROCm. Or more accurately when preparing to crash the system by attempting to use ROCm. Although in fairness as the other commenter notes it is probably a VRAM issue so I've been starting to suspect maybe the real culprit might be X [0]. But it presumably doesn't happen with CUDA and it is a major blocker to using their platform for casual things like multiplying matricies.
But if CPU and GPU share a memory space or it happens automatically behind the scenes, then the problem neatly disappears. I'd imagine that was what AMD was aiming for and why they tolerated the low quality of the experience to start with in ROCm.
[0] I really don't know, there is a lot going on and I'm not sure what tools I'm supposed to be using to debug that sort of locked system. Might be the drivers, might be X responding really badly to some sort of OOM. I lean towards it being a driver bug.
I have had a RX580, RX590, 6600XT, and 7900XT using Linux with minimal issues. My partner has had a RX590, 7700XT on Windows and she's had so many issues it's infuriating.
They might even be right. One of the potential advantages of the APU approach is if they GPU can be absorbed into the CPU with shared memory, a lot of the memory management of CUDA would be obsoleted and it becomes not that interesting any more. AMD are competent, they just have sucky crash-prone GPU drivers.