K20c and GeForce 9800 GTX on Ubuntu 14.04

I am having a very difficult time getting this setup up and running… I have tried following the instructions at: http://labs.beatcraft.com/en/index.php?CUDA6.5%2FUbuntu14.04#xdb14ad4 but everytime I complete them and reboot I get stuck in an endless login cycle (thanks lightdm) with extremely poor resolution. Does anyone have any hints? I am happy to provide more info as I realize this is probably not nearly enough.

If I “sudo apt-get purge nvidia*” and “sudo apt-get install nvidia-current” it tries to give the k20c the nvidia-302 driver and it seems that nouveau is still trying to bully its way back into running the GeForce 9800 GTX :( Thanks in advance!

Did you mean GTX 980? The GeForce 9800 GTX was a GPU with compute capability 1.1 (i.e., sm_11), and the most recent CUDA drivers do not support sm_1x devices any more. I cannot tell how recent the “302” driver package is, and whether it includes support for sm_1x devices. You may want to check the driver package release notes.

I mean the GeForce 9800 GTX. To be clear, I only need this device in order to drive my monitor, it does not need to do any GPGPU computing. I want to use only the K20c for CUDA.

Nevertheless you will need to choose a driver that is compatible with the 9800GTX, which restricts you to a 340.xx or 343.xx driver at the latest. This will restrict you to a CUDA 6.5 toolkit at the latest (which you already seem to be aware of.)

My suggestion would be as follows:

  1. Remove the K20c from the system.

  2. Do a clean load of Ubuntu 14.04. Make sure the GUI is working properly with the monitor attached to your 9800GTX. If it is not, this has nothing to do with NVIDIA software, since you haven’t loaded any yet.

  3. Stop the lightdm GUI and X, and switch to command line mode:

http://askubuntu.com/questions/320666/how-do-i-kill-the-x-server

  1. Remove the nouveau driver as described in section 4.3.5 here:

http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#runfile-nouveau-ubuntu

  1. Select an appropriate 340.xx or 343.xx driver for your system using the runfile installer, such as 340.93 from here:

http://www.nvidia.com/download/driverResults.aspx/91217/en-us

(you may want to download this runfile installer before stopping the GUI in step 3)

  1. Install the driver using the runfile installer. If you are prompted as to whether you want to install openGL libraries, select yes. If you are prompted whether you want to modify the xorg.conf, select yes.

  2. Restart the GUI (see step 3). At this point, the GUI should be working correctly. If not, stop and troubleshoot this before the remaining steps. You may want to test a reboot cycle.

  3. Power down the system, and install the K20c.

  4. Power up again. Things should still be working correctly. If the GUI is not working correctly, stop and troubleshoot this before the remaining steps. If the GUI was working correctly in step 7, but not now, it may be necessary to edit the xorg.conf file to identify the specific GPU to be used for X with a PCI indicator as described here:

http://nvidia.custhelp.com/app/answers/detail/a_id/3029/~/using-cuda-and-x

  1. Download the 64-bit CUDA runfile installer for Ubuntu 14.04 for CUDA 6.5 from here:

https://developer.nvidia.com/cuda-toolkit-65

  1. Run the CUDA 6.5 runfile installer. When prompted, select “no” to skip the installation of the bundled driver. Your previously installed 340.93 driver will work in this scenario.

  2. Apart from the above, review and follow the steps for CUDA runfile installation in the linux getting started guide:

http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#abstract

Thank you so much, I will give this a try. If I wanted to use CUDA 7.5 is there a budget GPU you would suggest instead? I was considering getting a NVIDIA Geforce GT730 as mentioned in the article, but it may have the same problems…

Will I notice any major performance hits from using the K20c with 6.5 etc?

How big is your budget? A GTX 950 (sm_52) costs around $150, and there is a whole line-up of increasingly powerful and expensive cards beyond that. If you just need a second device to drive your operating system GUI, a GT 730 may well be sufficient at roughly half that cost. Note that I have no hands-on experience with either card. NVIDIA website shows three versions of the GT 730, so you may need to do research as to which of these suits you best: http://www.geforce.com/hardware/desktop-gpus/geforce-gt-730/specifications. These GPUs are based on either the Fermi or the Kepler architecture, thus fully supported by CUDA 7.5 and current graphics drivers: https://en.wikipedia.org/wiki/GeForce_700_series

Thanks for your reply, I ended up getting a GT 730 DDR3, 64-bit for about $60 from Amazon. I figure the situation with the GeForce 9800 GTX can only get worse with future CUDA updates.

I have no real need for anything more powerful than the 730. If I need another GPU for computing I have another Tesla lying around.

The GT 730 should be here tomorrow. I will update this thread with any idiosyncrasies I discover during the install.