A computer

Gigabyte 890FXA-UD5
AMD PhenomII 1075T
2 x GTX-470
4 x Patriot DIMM 4 GB 1600MHx PC3-12800
Debian wheezy amd64

with

nvidia driver 275.09.07-1 compiled with Debian for linux-headers-2.6.38-2

is used for

(a) number crunching or - alternatively -

(b)graphic examination of results with CUDA-enabled viewer (VMD).

I am getting problems with both tasks. For (a), the best I found was from the linux prompt (without calling the X server)

# nvidia-smi -L

# nvidia-smi -pm 1

$ launch the parallelized code

$ monitor what happens from a ssh-linked desktop (the terminal of the gtx470 machine is as if it were hanged).

When completed (one-two days, a 20-100 GB output file), reboot from the shh-linked desktop and launch the graphic program, orking on the gtx470 computer itself. VMD creates its CUDA environment with the cards and normally it worked fine (it could not be launched without first rebooting).

Possibly after update/upgrade, VMD launching now hangs the system.

I have tried in alternative the desktop xfce4 with same problems. Removed xfce4. Then KDE, which was the worst and was removed, hopefully (not sure) removing everything with

# apt-get --purge remove kde*
# apt-get autoremove

Reinstalled gnome2.

Is anything peculiar about the X system and gtx-470 cards or there may be some wrong setting from my own?

thanks

chiendarret

For one thing, that is a Crossfire board and I don't believe that the chipset AMD 890FX/SB850 supports dual Nvidia cards.

I relied on the experience of the vendor, specialized in gaming computers.

If really only one card is supported, could you advise if a gtx-580 is supported by that mainboard? Or could you suggest an alternative mainboard to support the hardware that I posted ?

AMD PhenomII 1075T
2 x GTX-470
4 x Patriot DIMM 4 GB 1600MHx PC3-12800
********************

What I can say is that the simulation is severely demanding, based on a truly CUDA-parallelized code (NAMD) dealing with a system of over 250,000 atoms in my case. This means that the molecular dynamics work is spread over the cpu/gpu components (it does not deal of a simple launching of separate jobs) and it would hardly work if there is a hardware problem.

During the simulation, nvidia-smi queries revealed both cards and their temperature (not their % involvement, which probably could only be revealed with tesla cards), while "top -i" revealed all six cpus at work.

thanks

I must add that

nvidia-smi -q -d MEMORY

reports mem% usage for both gtx-470 while carrying out the simulation. In addition, for these simulation, the ratio cpu/gpu 6:2 is correct (must be at least 2:1). cpus carry out only energy calculations, most of the laod is to the gpus.

Also, before current nvidia driver 275.09.01-1, I worked for a week with 270.41.19-1 without any warning.

I asked to the debian amd64 site if problems were encountered on upgrading to 275.09.01-1 without getting any answer.

chiendarret

CUDA isnt SLI. You dont need SLI support for CUDA to work with multipule video cards. BTW, the SLI patch has been known to work with 890FX chipsets.

That makes me think its just an issue with the 275.09.01-1 driver, not your hardware. You could try getting the drivers from 'sid' (currently 275.09.07-5), or maybe even experimental if you were game. Or you could manually install 270.41.19 drivers. There might even be aq way to rollback to the 270.41.19-1 drivers as well, but I dont know how.....if its even possible.

I was wondering if your built that system chiendarret, sorry to hear its not playing nicely with the newest debian 'testing' drivers.

Please see also my thread "automount in xfce4 or gnome2" for the same computer. There is information that may closely relevant to the problems raised here for the gtx-470 cards and they are not in line with problems with the driver. Possibly there is something else that was not yet grasped.

Separately, I have asked to debian about how to safely install the nvidia driver from "unstable" or recover the older driver. Hope they will answer.

chiendarret

With the new nvidia-kernel-dkms 275.09.07-5 in Debian GNU-Linux amd64 wheezy, all described problems vanished. Thus, the suspicion that problems (absent with driver 270.41.19-1) were introduced by driver 275.09.07-1 was correct.
chiendarret