Our organization has an “official builds” group that produces all the officially compiled and packaged products. The systems these builds are performed on will probably not have NVidia cards installed. (These systems don’t need to run the software, just build it.) The systems have recent compilers installed but the hardware is be somewhat old, and the build group can be resistant to changing their hardware.
In the CUDA documentation, the first step to verify the presence of a compatible card. This is very likely to cause problems with the “build group”. Is there an official (or at least ‘common’) method for installing CUDA - for compilation only - onto systems that do not have an NVidia card installed?
yes. On linux, and assuming you don’t need applications that depend on the CUDA driver API, it’s simply a matter of running the linux runfile installer, and deselecting the option to install the driver.
get your runfile installers at http://www.nvidia.com/getcuda
follow the instructions in the linux install guide: http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html
On Linux, specifically RHEL 7, we are using the repo to install and manage the packages. What packages are needed to enable compiling of CUDA 8.0 and CUDA 9 programs without having a video card or NVIDIA driver installed?
It should be sufficient to install the cuda-toolkit-X.Y package.
Refer to the linux install guide for package manager install options: