So I am finally going to break down and buy a GPU for my desktop. Budget is very tight ($300 or less), so my options are pretty limited. I am considering the ARC B-series, RTX 5050/5060, and RX 9060XT. The Nvidia and AMD cards are 8GB of vram. The Intel cards are 10GB or 12GB.
This is for darktable image processing, I dont game and dont use LLMs, but may include AI denoise when the OSS stuff matures a little.
Phoronix sometimes includes darktable in their comparison articles, but it is hit or miss. I have looked into general OpenCL benchmarks but, tbh, I am not sure which are the most relevant to darktable’s modules. I am assuming FP 16 or 32? That is mere speculation on my part…
I have read here, and elsewhere, that AMD and Intel had driver issues in some instances when they came out. Do I still need to worry about that 1+ year later?
As a quick test I enabled openCL and configured it to use 8GB of system memory. I ran a few tests and did not see any tiling notifications in the logs (they do show up unless I force the 8GB memory use). So, 8GB of GPU memory should be ok…? Honestly, I have no idea how to gauge requirements here. I know the general idea is more is better, but that is only if all else is equal…and I dont know if all else is equal with these cards re: performance. Some of the benchmarks I see show big differences!
Am I over thinking this? Is real world performance probably similar enough that it wont matter which card I get…barring issues with getting openCL to be recognized?
In case it matters: Using dt_5.4 from my distro, or the build service now, so no flatpack concerns. I am running Deb 13-Stable with 6.16 kernel, and XFCE desktop.