titan x overclocking and rt gpu

[super gnu]

finally got my titan x plugged in, and watercooled.

works fine, using the old RT bench, i get a time of 1min 20 secs, when using it in conjunction with my old gtx 670.

temps are fairly static at 29-32 degrees, under idle or load.

as you might imagine i was quite excited to overclock the thing, since i heard with watercooling a 30-40% overclock isnt unheard of. im even happy to do bios mods to increase available voltage.


first attempt is, to say the least, baffling.

using evga precision x, im able to crank the clock from 1070 up to 1480 without any voltage increases. surprisingly high. i didnt try any higher yet.

however,

time in rt benchmark goes from 1min 20 to err.. 1min 19.5


hardly the result i was after.


i thought maybe there was some throttling going on, but in gpu-z i can see the clock is pegged at the new higher value, power consumption of the card is only 68% tdp (limit is 110%) temp is 33-34 degrees.

ive tried boosting the card voltage by 100mv, which increases the power usage to 75% tdp, but nothing changes.

what the heck is limiting me? surely i should see a decent speedup?

[savage309]

Try comparing the results when only the Titan X is rendering.

[super gnu]

i did, no change when overclocking.

i get a speedup in furmark opengl benchmark, but no difference in rt.

i did read another thread where 3LP had similar issues and said that restarting sometimes fixed the issue, but ive tried that a few times, and got no change.

[super gnu]

ok some numbers: seems opencl responds much better than cuda.

-please note these are single runs, with a low paths/pixel of 200, as im quite busy

-also note, im currently unable to increase memory clock. evga precision has a slider to adjust but it doesnt seem to affect the clocks im getting.



default core clock 1189mhz

opencl

854.7 kpaths

92.1 secs

cuda

1725.4 kpaths

45.1 secs



1475.2 mhz +24%

opencl

1019 kpaths

77.4 secs +19.6%

cuda

1838 kpaths

43.3 secs +4%


1509.3 mhz +26%

opencl

1038kpaths

76.2 secs +21%

cuda

1804 kpaths

43.1 secs +4.6%

[super gnu]

ok so some further research and testing.. seems the issue is using PrecisionX for the gpu overclocking ( or msi afterburner or any other "gaming" oc software.

issue is that gpus have different power states, P0 being the highest performing/clocked one, which is used for games and benchmarking ( and seemingly opencl from above numbers although this contradicts what ive read)

cuda (or at least cuda in vrayRT, and supposedly opencl too) use power state P2, which is more conservatively clocked.. presumably to guarantee stability.


setting the oc in precisionx or afterburner only set the P0 clock states.

if you use "nvidia inspector" to set the clocks, you have access to the overclocking for each power state individually.

when i use this software, my overclocking is reflected immediately in my cuda results.

im just doing some longer renders to get a stable value, then ill post some results.


interestingly if i look at the power draw of the card, an opengl benchmark such as furmark maxes out the power consumption of my card (110% TDP), and severely limits the overclock achievable, however vray RT never uses more than 80% of the maximum available power draw, despite also maxing out the gpu usage. if im just going to use the card for RT, it seems i can clock it a good 300mhz higher than if i want it to run furmark without throttling.

im not sure if this is a sign that vray has room to work the card harder..

[3LP]

Hum, interesting all this OC thing.
I also noticed that using EVGA precision doesn't always reflect my OC to the render.

It seems I can't OC though the P2 using the inspector though as my slider is already at the maximum.


What is P2, P5 and P8 anyways? What do they represent?

Stan

[super gnu]

they are power states that the gpu uses.. p8 -12 being low power states (idle, 2d workload, video workload etc) and the lower number states have increasing clocks. p0 being the highest state of the card for gaming. p2 is a slightly lower clocked state, presumably to guarantee absolute stability for compute workloads.

wrt turning up the p2 memory clock, 2 things:


firstly, it will never go higher than the p0 state clock, so you need to set that first.. note they use different number for the setting. p0 is an offset, p2 is actual clock.. f.e. i set my ram to +510 offset in p0 (for starters) then apply. look in the main nvinspector panel what clock that gives you, then go to p2, and turn the slider up to the same or slightly lower. you may need to hit "unlock max" in that window to adjust up.

note the max you can set is 4000 mhz there... you can however edit the saved "clock shortcuts" and turn it up higher.

secondly, you cannot edit the clocks if there are any cuda apps open. close rt server to allow changing the clocks.

finally, and ive no idea why, the p2 core clock setting has no effect. the core clocks still come from the p0 setting.