How to free PCIe card from display duties? HOW TO assingn display exclusive to IGP and CUDA tasks on

Does anybody know how to run the display exclusively on the IGP allowing the PCIe video card(s) to run exclusively GPGPU applications? On all the hardware I met the IGP gets disabled once a PCIe card is inserted or the add-in card becomes invisible if IGP is enabled. It would be nice to have display tasks decoupled from GPGPU crunching - no more “computing error” or display gone nuts. I tried once with IGP disabled and display allocated to an ATI 16MB PCI card but was not very happy (8600GTS was present in device manager but not in display manager).

To summarize:

1 - what chipset allow IGP display even if a PCIe video card is present. Is this related with the allocation policies of PCIe lanes?

2 - (if there are some) what manufacturer implemented this facility in hardware and let the user decide in BIOS

3 - what about compatibilities: nVidia cards on AMD motherboards, ATi on nVidia or any of them on Intel motherboards.

4 - how to install nVidia CUDA driver WITHOUT installing nVidia display driver? BTW I experienced “no CUDA device found. exit install” if a CUDA PCIe card was present (together a PCI card) but not as primary display. Must I plug a dummy display connector (which I dont have) into the video output of the CUDA exclusive PCIe card to fool it?

As long as TESLA exists the solution exist. But what is it? I need a platform with 1-2-3 GPGPU PCIe cards and keep the display duties exclusively for the IGP.

1&2 .Some motherboards allow you to specify the primary graphics adapter, PCI, PCIe, or integrated. Otherwise it is up to the operating system. I know for a fact in linux you can select the primary graphics card in xorg.conf which can be any device that is visible to the system; you can even select different drivers for different cards. I do not have any experience with windows.

If there is some restriction, you could try going with a smaller PCI or 1x PCIe graphics card for display that might not take up too much room in your case.

  1. In my experience it doesn’t matter. I am running 1 280GTX and 2 9800GX2’s in an AMD 790FX system and all 5 GPUs are detected correctly.

  2. I don’t think you can, they are bundled together. For several releases of the nvidia driver, CUDA has come included. I think that if you use a recent NVIDIA GPU you will have to use a driver with CUDA enabled.

I have an integrated ATI GPU, and an 260 GTX. I have to set in bios that the integrated card is used for display. Still, in windows I have to extend the desktop to the 260GTX(although it is not actually connected to a monitor), so that windows does not disable it. (I’m using widows XP, and latest drivers for 260GTX)

Intel’s recent integrated GPU chipsets (or at least the G31, G33, Q35) seem to physically switch out the northbridge GPU when a VGA card is detected in the PCIe slot. In linux, the onboard GPU is usually enumerated on the PCI bus as device 2.0. Put a VGA card into the x16 slot, and the PCIe card gets enumerated as device 2:0 with the northbridge CPU nowhere to be seen. I certainly have been able to work out a way to make both coexist. I guess it is possible that the G45 is different, but I very much doubt it.

Easiest solution if your building a new PC is probably to get a MB with NVIDIA chipset in the first place. I got a MB with Nvidia’s 8200, and plugged in a GTX260, the only thing I had to do to get CUDA to run on the 260 without a watchdog timer was enable the onboard gpu in the BIOS. After installing drivers (in XP x64) and without extending the desktop onto the 260, it ran fine.


I have a 9600GT with my M3A78-EM motherboard. The MB has video on, and the chipset is HD3200. if i use the 9600GT only CUDA, connecting the monitor in the MB, and use the HD3200 for video everything would be fine, right? and if i use the 9600GT AND the HD3200 to program (the ATI site says that is possible, but the chipset is very useless when it is compared with the 9600 - no double precision, no kernel scatter), and use one of those (the chipset) to video, there will be any problem? Will be some problem, using 2 GPUs to run programs at the same time? suppose that the chipset is a nvidia, to not talk on ATI if was preferred.