Octane 3.0 and GPU Extenders - ISSUES ?

[Lewis]

Hi!

I'm testing new system and i have some problems with octane 3.03 speed when rendering with GPUs on extender (1x sped). I KNOW that uploading files to GPUs on 1x extender has to be slower and i understand why, part what's not clear to me is WHY it renders noticeably slower then on Motherboard. I had no such problems with octane 2.x

I have system with 4 GPUs now (so i had to put some to extender just to make them have space for air in between so what i did was put 1 GPU on PCI-E 1 and 1 GPU on PCIE-E 7 both directly on Motherboard and then 2 more GPUs on Extenders 1x at slots 3 and 5.

Simple scene testing for PR kernel 2500 samples amd Resluts are like this:

- TitanX (MB x16) - 134 sec
- GTX 1080 (MB x8) - 145 sec

- TitanX (MB x16) + 1080 (MB 8x) - 71 sec
- TitanX (MB x16) + 1080 (extender 1x) - 82 sec (this is 15% slower already when on extender)

- TitanX (MB x16) + 1080 (MB 8x) + 1080 (Extender 1x) - 56 sec (this is also only 15 sec faster than TitanX+1080 on MB which is noticeably slower than 15% slowdown)

And on top of that, as soon as i enable any GPU on extenders my plugin (LW verison) windows are behaving very sluggish and slow refresh (IPR priority is set to low and tonemap set only to 1 GPU in MotherBoard at time so it's not that). As soon as i enable GTX 780 (that's another one i have in extender 1x but didn't include in this test it 'coz it's slower) then everything goes another level of slowness and GUI refreshes problems (even if display GPU is off form rendering).

So do we have some GPU mix problems (Kepler, Maxwel and Pascal in same case), driver issues (latest nvidia 372.xx) or Octane is doing some weird massive uploads to GPUs constantly and making it very problematic on extender GPUs ? Octane 2.25 works great with similar setup (i had 5 GPUs and 2 were on extenders) and no problems, no tonemap issues, no slowdown, GUI is snappy and responsive....

So is there way for otoy DEVs to make it stop doing so much PCI-E traffic and enable GPU only mode (like 2.x was for frame buffer GPU only) so that users can choose FAST mode with VRAM only or use PC RAM if needed for bigger images/fram b uffer ?

Please OTOY can you try to do something, you knew lot of people use extenders (nobody is "happy" with just 2-3 GPUs and stacking 4 of them in MB is too hot/overheat due being too close - i tried that and throttling of GPUs was regular with 3rd party GPUs so i opted for slots 1,4,7 and extender sin between but i plan to have extenders in most slots to get 6 GPUs working in 1 case/MB)

thanks

[glimpse]

Hi Lewis.

Just an idea to check.

try running octaneBench & fireing GPUZ in the background. Might be a case where GPU starts to be limited by other factors, like for instance x1 isn't able to provide enough power..-without graphs (monitoring vital GPU functions) it's impossible to tell exactly what's hapenning in Your system. I'm not to say Octane's code has nothing to do with that, but Your configuration might also make a bit of a difference.

[Lewis]

Hi Lewis.

Just an idea to check.

try running octaneBench & fireing GPUZ in the background. Might be a case where GPU starts to be limited by other factors, like for instance x1 isn't able to provide enough power..-without graphs (monitoring vital GPU functions) it's impossible to tell exactly what's hapenning in Your system. I'm not to say Octane's code has nothing to do with that, but Your configuration might also make a bit of a difference.

Hey Tom!

Thanks for idea but I have/had nview monitoring of GPUs to see temps, clocks, RAM usage and all so it's not that. it just slow to have it on 1x nowdays at octane 3.x .

[jang2235]

Ver 3.0 released shortly after I was aware of the problem.
Contact the result, but the result was not the solution. In Ver 2.0 did not occur, the problem appears only in version 3.0.
I will only do wrong can not understand how it can make a significant performance degradation. This does really judge the slow data transfer faster than gigabit Internet speed problems are ahningeot it.
Sounds like not to use the Rig.

[mediafx]

Hey there, got any answers from the guys @otoy?

I've been experiencing the same performance issues issues

[Lewis]

Hey there, got any answers from the guys @otoy?

Nope, so far nothing form Otoy regarding that problem.

[preciousillusion]

Hi Lewis.

Just an idea to check.

try running octaneBench & fireing GPUZ in the background. Might be a case where GPU starts to be limited by other factors, like for instance x1 isn't able to provide enough power..-without graphs (monitoring vital GPU functions) it's impossible to tell exactly what's hapenning in Your system. I'm not to say Octane's code has nothing to do with that, but Your configuration might also make a bit of a difference.

If the 1080 used with the riser is the FE (reference) model using a single 8 pin power connector there could theoretically be a power issue; if the card is somehow not able to negotiate the full 75W from the pcie-slot when connected via the riser (the first 11 pins up to the notch are for power, presence detection etc), and thus only gets the 25W (8 pin 150W + 25W = 175W - 1080 FE TDP 180W). If the card’s got 8 + 6 pin, there’d be at least 250W.
But since there's no problem with V2 it would also require that V3 somehow go bananas specifically on Pascal power consumption as on 970, 980, 980ti and Titan X (the ones I've tested) it's usually around 40% of TDP, never exceeding 50%.

It's definitely not a bandwidth issue (as long as “out of core textures” isn't used, and probably not even if used as there's 1 GB/s bandwidth on a pcie 3 x1) as the scene has to be loaded in gpu memory before rendering and the only significant traffic on the bus is an initial burst when the scene is loaded.