9800 GX2 & Tesla problem Which driver & CUDA version?

motherboard XFX 790i Ultra 3-way SLI
9800 GX2 in the first PCIx slot
tesla C870 in the second PCIx slot
Intel Q9300 quadcore
CoolerMaster 1000W PSU

Windows XP 64 Profesional

9800 GX2 and Tesla C870 are not connected via Sli bridge

After installing 169.21 drivers 9800 GX2 is not recognized and GL window inside Nvidia Control Panel is empty. I uninstall that driver and try with 174.74. After that windows recognized both 9800 cores and GL works fine but now Tesla is not recognized. Why?

Is it possible to make all three GPUs work and how? Which CUDA version should be installed to support all of them and which nVidia driver version?

Since one PCIx 16 is still free can I use for example an extra GeForce 6800 card as display adapter and have all 3 GPUs available for CUDA?

Can’t answer the other questions, but you can’t run CUDA programs on the GeForce 6800. Only GeForce 8, 9, and Tesla have the GPU architecture required to support CUDA. You could still use the 6800 as display-only, though.

I know that. The idea was using 8600 as display adapter only and 9800 gx2 (has two 8800 GPUs as I know) and Tesla core for CUDA. Which driver can support that?

Also in CUDA documentation 8800 ultra and 8800 gtx are reported as 1.0 architecture and 8800 GTS as 1.1 architecture but what about 9800 gx2?

It contains two 8800 GPUs but which architecture?

Tesla is 1.0 architecture. Does CUDA support mixing of different core versions?

9800gx2 is G92-based, so it’s 1.1 arch.

Basically you cannot run 1.1 code on 1.0 device. In addition, I saw it somewhere said you cannot mix the compiliation of 1.0 and 1.1 together, but…I think it should be okay as long as you send the right code to the right GPU.

By the way, I don’t know if you can use 8600 in the third PCI-e slot as the main display. I think you will have blank screen while booting because BIOS would assume screens are attached to the first available GPU, which is 9800GX2 in your case. I dont know if that just my mobo (ASUS Striker Extreme) problem or a general problem for all mobos.

To me this is not a big deal since I am using Linux, and I can always shutdown x server to avoid the 5 secs limit and some instability problem, but it’s simply not possible under Windows.

Is there anybody who successfully use the third or the second slot (while the first slot is also occupied) as the primary display with boot screen?

If display adapter is forced to be in slot1 I can put 6800 in that slot and 9800 gx2 and Tesla in slots 2 and 3. But main problem still exists. Which driver can recognise both Tesla and 9800 GX2 cards?

CUDA display driver should be able to recogize all of them, including 6800 because of the unified driver architecture. The drawback to install 9800GX2 on the second or third slot is lower bandwidth. Because in my experience, the first slot usually has higher bandwidth.

Download 2.0b driver here (2.0b would recognize 9800GX2):


As Mu-Chi Sung wrote (thanks!), the beta 2.0 driver is the best place to go. The 169.21 driver is too old to know about the new 9800gx2 card. And if you look at the description of the 9800gx2-specific driver (174.74) it doesn’t support Tesla yet. We tend to do this with new cards, just support that card alone initially in it’s specific driver.

the beta 2.0 driver should support both of these.

Hope that helps. I’m not sure how far I’d try to get that 6800 going for now. Not sure it gets you much since you’ll be able to do plenty with the 9800gx2 and C870. I’d start with just those 2 initially and the beta 2.0 driver. Then add the 6800 later.

Thanx, I will test it today. 6800 was an alternate option, but primary goal is to make 9800 gx2 and tesla to work togather. It seems it is possible with 2.0b driver. I will post results after testing.

Thanks again

another update for you: today we posted a released driver for XP/XP64 that supports this combination. See in the standard driver downloads page http://www.nvidia.com/Download/index.aspx for 9800GX2 (or GTX) the 175.16 driver. Note this is for use with CUDA 1.1 on XP, not Vista.

I got problems in both drivers 174.55 and 175.16 on CUDA 2.0 on XP, when I was executing the project - volumeRender. In 174.55 it could work but after that I got a blue screen of death which showed me "C000021a unknow hard error. " So I reset the PC, reinstalled the driver to 175.16. I could execute the volumeRender with 175.16 (still with CUDA 2.0 on XP), but had no rendering coming out. Does that mean 175.16 for use with CUDA 1.1 on XP only? How can my problems be solved? Thanks.

My PC spec is 8600GTS, Intel Q6600 and 32-bit XP.