I think that NVIDIA is using a ridiculous driver naming convention.
If I look for a Tesla driver download for Linux/AMD64, I get a file called NVIDIA-Linux-x86_64-331.89.run, and the “SUPPORTED PRODUCTS” tab lists Tesla cards.
If I look for a GTX driver download, I get a file called NVIDIA-Linux-x86_64-340.32.run, and the “SUPPORTED PRODUCTS” tab lists GTX and Quadro cards.
Since I have both cards, and I am trying to get CUDA 6.5 working with them, should I install both drivers at the same time? Or are these different (and presumably conflicting) versions of the same driver?
You should not install both drivers at the same time (I don’t think this is really possible, anyway).
Since you are on 64 bit linux, I would just use the 340.29 driver that comes bundled with the CUDA 6.5 64-bit linux runfile installer. It will work on all Tesla, Quadro, and GeForce cards that support CUDA, with the possible exception of cc 1.0 cards that are no longer supported by CUDA 6.5 (or current drivers).
The 340.32 driver would likely also work as well.
The 331.89 driver will not work with CUDA 6.5
If you want the 340.29 driver as a standalone installer, you can extract it from the bundled CUDA 6.5 installer using some command line switches. Launch the CUDA 6.5 runfile installer with the --help command line switch, and it will be clear.
Thanks for the suggestion! I’ll probably try that first. Just to help me understand NVIDIA driver versioning better: Can different driver versions coexist on the same system?
The CUDA download FAQ says:
This seems to imply that it’s OK to have different driver versions installed. Will both of them be used at the same time (especially if they support non-overlapping sets of devices)?
It’s not possible as far as I know to have 2 NVIDIA linux drivers installed and active on the same system at the same time.
I think you are misreading the text you excerpted.
The bundled installer includes an option to install the included (bundled) driver. IF YOU SELECT that option to install the bundled driver, the driver installer WILL REPLACE the driver that is currently installed on your system. It is not giving you the option to have both drivers. The only way the nvidia linux driver installer will proceed under ordinary circumstances, AFAIK, is if you agree to remove the old driver if it finds one. If you do not agree to remove the old driver, the new driver install process will quit without installing the new driver.
If you don’t select the option to install the bundled driver, and your existing driver is not compatible with CUDA 6.5 (basically r340 or newer) you’re going to have problems when you try to run any CUDA code on your system compiled with CUDA 6.5.