GeForce + Radeon => is it possible?

Hello guys!

Recently I was considering the purchase of a new GeForce (for example 9800 or GTX280) for my CUDA computation. I have a computer with Radeon (PCI-e) graphic card, so I have some questions:

  1. Is it possible to put into my computer new GeForce card, and using both Radeon and GeForce at the same time? For example Radeon for displaying and GeForce for GPGPU.

  2. I know that computation on graphic card that is used also for displaying is a bit problematic because of watchdog stuff. But I don’t know if this apply to computers with more than one graphic card, for instance first GPU for displaying and second for GPGPU? I assume that graphic cards are non-Tesla (with DVI output).

PS. Sorry if such questions have already been discussed on this forum, but I couldn’t find them.

Won’t work on Vista, might be problematic on XP, should be fine on Linux.

Are you talking about Radeon and GeForce in one computer or about problems with watchdog?

I know that watchdog can be turned off in Linux, but can’t be turned off in Windows. My question was, if there are any problems with watchdog if you do computations on GPU (for example 9800GTX) that is not connected to a screen.

Both, basically!

Vista can’t handle drivers from different vendors, period. Linux will work fine if you’re not using the NV card from X (so initialize the device entries using the script in the release notes). XP might work with no watchdog timer, but it’s very picky about drivers, and installing two vendors’ drivers simultaneously has caused problems for me in the past.

Thanks for reply. Do you (or anyone) know if Windows XP turn off the watchdog for second graphic card automatically? Or maybe you have to do it yourself? I didn’t see such option in XP :/

There’s no watchdog timer on any GPU that is not using a display (as of 178.28), so it should be fine.

I have faced problems with an ATI card… If ATI is the primary display, NVIDIA’s drivers won’t even be loaded in Win XP – That was my understanding. your CUDA programs will NOT even work… It is one driver at a time… Thats my understanding.

In my PC, I have just disabled the ATI card and am using CUDA for computation and display. So 5-sec timer always hit me…

Not sure, if my card is ATI radeon though… Itis from ATI. Dats all I know.

But I was using my PC remotely by then… So, I could not really debug the problem… Disabling ATI worked and I stuck to it… But I was less CUDA aware by then…

The watchdog is not a problem for most kernels.

Also I don’t think you should hack around it because ultimately it is very useful. A CPU can run multiple programs at once, but a GPU can’t. I believe that on Windows, if your kernel goes into an infinite loop on a non-watchdogged card, you basically have to restart your computer. (Although on Linux I think you can just Ctrl+C.)

May b, one should just extend the desktop that GPU to kill that process and then shrink the desktop again :)

Not sure, if that would work though…

If not, nvidia should provide a tool to do this… Restarting the computer looks odddddddddd…

To sum up:

if I want to work with watchdog disabled on Windows, I have to have two nvidia’s graphic cards (not ATI), one for displaying and one for computation.

Do I understand it correctly?

(As for Linux everything is clear for me).

By the way, I’m curious if this two nvidia’s graphic card have to be the same or can I use for instance a 8600 (displaying) and a 9800 (GPGPU)?

You are absolutely right atleast for WinXP. The display card need to be supported by the NVIDIA CUDA driver (Check out the driver site… It will list a big list of cards that it supports). Select the cheapest and decent card for your display.

Use the high-end for your computation.

The good news is that NVIDIA CUDA drivers support many low end graphics card…

and, do NOT forget to disable your ATI card in Windows before installing your second GPGPU card (if you do it before installing the first cheap card then you wont have a display at all :) )