Is Tesla D870 compatible with Vista x64?

Hi…

Still struggling with Vista x64 here…
I have GeForce 8600GT + Tesla D870. Latest CUDA drivers installed from the website.
When I run deviceQuery.exe from the CUDA SDK it only detects the GeForce and doesn’t detect the Tesla. Debugging confirms that only one card is being detected here.
Things I’ve done:

  1. Reinstalled the driver. Didn’t work.
  2. The Tesla installation CD doesn’t work at all (Error message says that the installation is only for Windows XP).
  3. NVIDIA Control Panel -> System Information, detects all the cards properly.
  4. gpuZ - detects all the cards properly.

Any ideas?

Thanks!

Y.

I just found this:
http://forums.nvidia.com/index.php?showtopic=76537&hl=tesla

Where Sumit Gupta says:
“We are actively working on supporting Tesla under Vista. We expect full support within a few months.”.

Any updated estimate on this matter? It took me quite a while to set up this system here with Vista x64, it’s quite disappointing to realize that Tesla drivers do not work under Vista… The irony here is that Tesla is really useless under XP 32 bit because of physical address extension issues (Out of my RAM’s 4GB, I only have 1.5GB left with all the cards installed, I had a post about it in these forums somewhere). The memory issue is solved in x64 systems. Will I be forced to switch to XP x64? I’m not so enthusiastic about that…

I would love to know if there’s some kind of a roadmap for the Vista support so I could plan my next move here…

Thanks!

Y.

2.1 beta should add official Vista support, so it should be available in a matter of weeks. However, I think that XP x64 is still the best way to go.

Where do I sign up as an Alpha tester to get the first release…? :-).

I’m not inclined to go for the now really outdated XP. I think I will stay with Vista and just wait for the next CUDA driver… If it’s just a matter of weeks then I’m willing to wait…

Thanks…

Y.

There are non-trivial issues with Vista versus all other OSes, the most prominent being that for the immediate future all GPUs will have a watchdog timer enabled (that may be shorter than other OSes).

Just an FYI.

edit: okay this is not entirely true, apparently you can change the watchdog timer behavior via a registry key: http://www.microsoft.com/whdc/device/displ…dm_timeout.mspx

the driver team called me an idiot because I didn’t know about this!

First, have you tried it? It has been mentioned on the forums before, but I can’t think of a single case where someone said “yes, this works”.

Curiosity killed the cat, so I just had to try it out: Running CUDA 2.0 (178.08) on vista64 I wrote a test app that runs increasingly longer kernels. I tried changing TrdDelay to 0xA with no effect. I tried changing TrdLevel to 0 (off) with no effect. I even rebooted between changes just in case.

Second (assuming that the reg keys did work), for apps professional distributed to end users I don’t see it as a viable option. Microsoft lists these keys as for testing purposes only. And mucking about in the hardware registry settings is not something that I would want typical user software doing. Software that does that is usually clearly labeled as hardware tweaking software and comes with all sorts of warnings when you start it up.

Let’s not talk in generalities. But yes, the watchdog timer is a useful feature that increases system reliability. It’d be a shame if it had to be removed to use CUDA.

Plus, this is something I wondered about but never answered myself because I take the timer as a given – If a kernel is stuck in an infinite loop, how do you manually kill it?

The only way I know of is by resetting the computer…

In linux, Ctrl-C works great. I’ve never tried this in windows as I’ve never run on a system without a watchdog in windows.