Has anyone gotten a dual CUDA installation to work? We’d like to have both 2.3 and 3.0 versions running on the same Linux installation.
It seems that we’d have to have to graphics driver versions installed… is this correct? Is it true that the graphics driver version 190.x only works with CUDA 2.3, and that the new driver version 195.x only works with CUDA version 3.0?
How would it be possible then to have both graphics drivers installed, and to switch between them? Would it be possible to specify this with a startup flag, at boot time?
You should be able to run CUDA 2.3 with a 195 series driver. I have and use both 2.3 and 3.0 on boxes with 195 drivers and have never had any problems. If you want to preserve OpenCL support as well, you need to be a little more cautious about which drivers you run, but for CUDA it should be fine.
I think it should depends which NVCC compiler you use. You may overwrite the nvcc compiler when you install new version of CUDA. Try install them in different directories. Set you environment variable correct. You may compile your code with different version of nvcc compiler. Hope it works.
The Yellow Dog for CUDA has a CUDA switcher included. It supports using multiple versions of CUDA and OpenCL. You can easily use cudatoolkit 3.0, 2.3, and 2.2 on the same machine and easily switch between them using configuration files or environment variables.