Installing CUDA

The basic steps are as follows:

  • Ensure drivers are appropriately up-to-date (e.g matching the versions specified on the CUDA page);
  • Download and install CUDA Toolkit;
  • Download and install CUDA SDK.

Ubuntu Notes

  • The 256 range drivers can be added by including the x-swat ppa repository, and installing nvidia-current, nvidia-current-modaliases and nvidia-setting. 196 drivers are also available.
  • Ubuntu is normally at least one release ahead of the supported CUDA release. However, CUDA 3.1 works on 10.04 as well as the suggested 9.10.

Fedora Notes

  • 196 drivers are available from rpmfusion using this guide. However, CUDA 3.1 requires 25* range drivers, which have not been packaged as of yet. One option is to build your own package.
  • CUDA 3.1 supports Fedora 12 - the toolkit + SDK work fine on this release.
  • CUDA 3.0 requires GCC-4.3, Fedora 12 has GCC-4.4 and cannot be easily downgraded. There are workarounds for this on the NVIDIA Forums

Windows Notes

  • The CUDA page links to the development drivers, but the released drivers generally work fine and are more up-to-date. They can be found at Note: You _must_ use drivers in the 25* range if you plan to use CUDA 3.1.
  • The CUDA SDK uses Visual Studio to manage the projects. Since we depend on the SDK, Visual Studio should be installed under Windows. The eventual plan is to produce Visual Studio projects from the compiler, and use "msbuild" to compile them. For now, the generated code can simply be copied into a new Visual Studio project, and compiled with the appropriate dependencies on the SDK.
Last modified 10 years ago Last modified on Aug 12, 2010, 1:23:05 PM