We are still actively working on the spam issue.
Difference between revisions of "CUDA"
Latest revision as of 21:08, 17 June 2022
CUDA is the nvidia libs for GPU programming for accelerated matrix manipulation and a requisite for CUDNN, nvidias dedicated tensor core library.
Installation Guide (+ CUDNN)
Installing Nvidia CUDA is shit, to successfully install the full cuda sdk and cudnn on a modern gentoo box required a bit of patience, trial and error. So this short guide should help you avoid a repeat of past mistakes.
If you're running an older version of the gentoo kernel, or a legacy driver, be sure to have the driver installed, signed (for hardened users) and loaded, you may not get all the features of a modern dist, or parts of this tutorial may not apply. Use discretion.
I have not yet tried on a no-multilib system, so I cannot guarantee functionality or a successful compilation.
1. Make an nvidia developer account on the nvidia-developer portal.
2. Download the appropriate CUDNN binturd.tar.xz and copy to /var/cache/distfiles
3. You will need to change your compiler version to nvidia spec, which means an emerge <=gcc-11 (I went with gcc 10.3.1 as of the time of writing this.) Select with gcc-config.
4. Ensure you're using =nvidia-cuda-toolkit-11.5
5. update make.conf accept license var.
6. emerge nvidia-cuda-sdk, If you've already installed it, be sure to recompile with the correct gcc version.
7. Utilize the library; pytorch, tensorflow, etc.