AMD is getting better for ML/scientific computing very fast for the regular consumer GPUs. I have seen the pytorch performance more than double on my 6700xt in 6 months to the point that it has better performance than a 3060(not ti).
Comment on The Maintainer Of The NVIDIA Open-Source "Nouveau" Linux Kernel Driver Resigns
cbarrick@lemmy.world 1 year ago
The state of GPUs on Linux:
- Want to game: choose AMD,
- Want to do ML/scientific computing: choose NVIDIA,
- Want to do both: fuck you!
beigeoat@110010.win 1 year ago
csolisr@communities.azkware.net 1 year ago
Great to know, but unfortunately CUDA is still too entrenched in comparison to OpenCL - hopefully there’s some way to convert from one to the other.
cbarrick@lemmy.world 1 year ago
OpenCL is dead. Vulkan is the future.
We still need core libraries to be written as Vulkan compute shaders, similar to cuFFT and cuDNN.
There is VkFFT, but I don’t know how good it is. I don’t think anyone has written a VkDNN library.
NVIDIA has signaled that they will fully support Vulkan.
AdrianTheFrog@lemmy.world 1 year ago
What about SYCL?
Asudox@lemmy.world 1 year ago
There’s also ROCm from AMD that is suppoed to be the rival of CUDA.
youRFate@feddit.de 1 year ago
Gaming has been perfectly fine, just use the proprietary driver.
Edgelord_Of_Tomorrow@lemmy.world 1 year ago
Congratulations you just killed 2/3 of all Linux users reading this post
youRFate@feddit.de 1 year ago
I mean, I play proprietary games, no need to insist on foss drivers…
BetaDoggo_@lemmy.world 1 year ago
Nvidia can game just as well as it does on windows, proprietary driver inconveniences aside.
woodgen@lemm.ee 1 year ago
I’m doing both on AMD, while I contribute to a bunch of ROCm packages. Stable Diffusion XL and training runs great.
csolisr@communities.azkware.net 1 year ago
Or buy a motherboard with dual slots and install one of each, hopefully you’re made of money