Workstation, Folding Rig, and Photos

[ccokeman]

Nice! 

[scr4wl]

Having some weird issues with the cards.

 

Part of the issue comes from me trying to run this headless.

 

I want the fans to spin at 100%, unfortunately nvidia-settings does not work unless you have an xserver running.., that's dumb.

 

The work around is to start an xserver with a dummy monitor on each card, adjust the fan speeds with nvidia-settings, and them kill the xserver.

 

Not a big deal and I've already set up a script that does all this automagically.

 

Here is the problem though. When I set the fan speeds to 100% (or manually set them to anything as far as I can tell) some of the gpu's are throttled.

 

GPU0 always works fine, the fan can run at 100% while maintaining mem and core clocks.

 

Everytime at least one gpu does not work appropriately. The clocks (according to nvidia-smi) drop down to 600mhz and the power usage drops from 250 to 30W. Utilization stays at 100%.

 

Sometimes this only happens to one card, sometimes 2, sometimes 3,

 

I have not touched the power limit.

 

Also, I cannot manually readjust the clock speeds. I'm using 12 as the bitmask for coolbits, which i thought was supposed to give me fan and oc control, but I can only adjust the fan. (I've also tried 5 and 8)

 

Any ideas? I'm using newest drivers and just want to run the fans at 100% without loosing performance.

 

Edit:

 

Also, even if I only adjust the fan speed of gpu0, some of the other cards are throttled.

Edited July 5, 2017 by scr4wl

[ir_cow]

I'm guessing you are using Linux so you cant use the Gigabyte software?

[ARandomOWL]

Everytime at least one gpu does not work appropriately. The clocks (according to nvidia-smi) drop down to 600mhz and the power usage drops from 250 to 30W. Utilization stays at 100%.

 

I don't have a proper solution, but you could try forcing the application clock with nvidia-smi.

[scr4wl]

I don't have a proper solution, but you could try forcing the application clock with nvidia-smi.

I've tried doing this, but I get an error saying that changing the clocks is not supported for the

Gpu's although I know it is, because it says so in the driver release. Also using 12 as the bitmask.., maybe that's not correct but i thought it was. I may try an older driver version.

 

If I just use two cards I can get both fans to spin 100% without throttling.

 

I'm going to try with three.

 

Also @ir_cow, yes debian

 

Edit: with three cards I was able to increase the fan speed on 2 cards to 100%, not a third card though. If I kept one on auto though, the other two would be fine.

 

 

With four cards though at least one is throttled. Look at consumption of gpu1 while at 100%

 

Edited July 5, 2017 by scr4wl

[ARandomOWL]

 

I don't have a proper solution, but you could try forcing the application clock with nvidia-smi.

I've tried doing this, but I get an error saying that changing the clocks is not supported for the

Gpu's although I know it is, because it says so in the driver release.

I've done this successfully on a 980Ti on 381.22 Not sure why fan speeds can't be set through nvidia-smi I leave my fans on auto for now.

[scr4wl]

I've done this successfully on a 980Ti on 381.22 Not sure why fan speeds can't be set through nvidia-smi I leave my fans on auto for now.

Looks like I'll be doing the same. I strapped some extra fans into the case blowing directly over the two cards inside of it. Sorta helps.

 

I dont understand why increasing the fan speed causes some of the cards to throttle though lol,

[scr4wl]

So photobucket wants to charge a whole bunch for image hosting now lol.

 

Anyway, I added a 1080 to this mess (sorry for blurry photo)

 

 

 

I was also able to figure out the fan throttling problem.., well sort of. It turns out that the cards enter a low power state as soon as x is closed (but only after running nvidia-settings ..., weird)

 

Anyway, the work around is pretty simple. You just need to throw your nvidia-settings commands into a script that starts after x starts (or just start x in the same script first) and then keep x running indefinitely with something like an empty cat command.

 

Now I can run the fans at 100% at the GPUs don't enter the low power state.