GeForce 9 & Linux support status

Accelerated Computing CUDA CUDA Programming and Performance
1

Can anyone shed any light on the support of the following configuration:

RHEL Linux 4, x86_64
CUDA 2.0
GeForce 9800

Supported configs don’t seem to be well documented on the Nvidia site. If you go to the driver page, there is no Linux support listed at all for GeForce 9. However, there appears to be a beta 173.08 driver with the needed support:

[url=“Page Not Found | NVIDIA”]http://www.nvidia.com/object/linux_display_ia32_173.08.html[/url]

A more recent driver is available from the CUDA 2.0 download page (175.55). However, the release notes for this driver don’t list the GeForce 9.

Thanks for any guidance anyone can provide.

cheers - jc

2

All released hardware is supported in the CUDA_2.0-beta packages.

3

Is it supported in Linux x86_64?

Which driver should be used for the GeForce 9 series under Linux with CUDA 2.0?

thanks - jc

4

Please read it in the forum annoucement. Everything you need is right here:

http://forums.nvidia.com/index.php?showtopic=65067

But I think currently 2.0b does not support 9800GTX officially. (9800GX2 is supported though)

http://forums.nvidia.com/index.php?showtopic=64997

5

I can confirm, I’m currently running CUDA 2.0b on CentOS 5 with a 9800GTX.

CUDA works alsmost perfectly without any tweaking involved (only doesn’t recognise the full card name and reports the code name “G92-420” instead of the full product name “9800 GTX” like previous versions did).

Note that my server is text-mode only (only a plain simple PCI card running a text terminal just in case we need some local access). There’s no Xserver at all and the GeForce isn’t connected to a monitor (to avoid the execution time limit, and in general to avoid any external disturbance to the running kernel that may arise from a concurrent graphic display), so I can’t assert if 3D is working with that hardware/driver combination.

Also the nice thing in this configuration is that the 2.0b SDK has separate libraries without (libcutil) and with(libparamgl) dependencies on Xorg. (Whereas everything was in libcutil in previous versions, although it was still possible to obtain a non-Xorg-dependent libcutil). Which makes it easier to compile CUDA code on a machine without Xorg.