GPU benchmarks

[cray-zard]

heres mine.. benchmark 2.0

rendered in single 980ti.. 9:55

[super gnu]

[ATTACH=CONFIG]26048[/ATTACH]

heres mine.. benchmark 2.0

rendered in single 980ti.. 9:55

Wow that's faster than a titan X. Is it overclocked ?

[cray-zard]

nope its just a zotac 980ti ref..

[eyepiz]

4 x Titan X - 30s
[ATTACH=CONFIG]26019[/ATTACH]

edit:

8 X Titan X = 14s

Think maybe the scene is "too easy" to test with more power?

Odd...

I ran (4) Original Titans at 28.04 sec.

http://forums.chaosgroup.com/showthr...335#post579335

[AlexP]

Odd...

I ran (4) Original Titans at 28.04 sec.

http://forums.chaosgroup.com/showthr...335#post579335

Would different manufacturers make a difference? Mine was on Max 2015 also?
No idea

[auke]

Hmm, just did a test run here on a Titan X and only getting 18 min and 10 seconds. I also notice an average of 1800 kpaths/s. Seems like the performance is practically split in half looking at the graphs. Could this be caused by my rig running a Quadro (K4000) and Titan X at the same time? The Quadro was not rendering though and I'm running GTX drivers. Running it all on 3ds Max 2016 and Vray RT 3.20.02.

[3LP]

7 x Titan X OC - Benchmark 2.0 : 1m33.4 (did it twice to be sure )
I wanted to get a rig that was equivalent but cheaper than the VCA, seems the challenge is done



Stan

[AlexP]

Nice

All on one board?

[3LP]

Not yet, I'm still working on it
I ordered a new riser that should fix my issue as it's the best one available, should be there next week, and it we will see.
If it doesn't then I don't know how to make it work beside changing the motherboard, and man, I hate Asus boards, I should have gone for Asrock, those never let me down.

Stan

[auke]

Keep us updated! Very curious how it will work out, also what kind of trouble are you having with the ASUS board (and which one are you using)?

Also a question for those using maxwell (gtz 9xx, Titan-X or M6000) cards. Could you please check what P-state your gpu's are in while rendering? It seems mine is stuck at P2 which halves the memory clockspeed, meanwhile I get benchmarks 40% lower than what I should expect looking at the benchmarks here.

[AlexP]

Keep us updated! Very curious how it will work out, also what kind of trouble are you having with the ASUS board (and which one are you using)?

Also a question for those using maxwell (gtz 9xx, Titan-X or M6000) cards. Could you please check what P-state your gpu's are in while rendering? It seems mine is stuck at P2 which halves the memory clockspeed, meanwhile I get benchmarks 40% lower than what I should expect looking at the benchmarks here.

How do I find out what P-state?

[super gnu]

Cuda tasks always use P2 power state on maxwell. its a driver "feature" to guarantee stability for hpc workloads. ive asked Chaos if there is a way to get around it. thats not the issue for you , as all the maxwell cards have the same handicap.. i dunno about the other chips.

having said that, you say your memory clock is cut in half? Thats much more extreme than ive found on my titan x (bog standard "cheapo" gainward) . mine goes from 3300 mhz in P0 state, down to 3000 mhz in P2.. hardly the end of the world.

you can easily get around any clock throttling with NVinspector. set your P0 and P2 memory clock speeds to the same value (4000 mhz in my case) and watch vray fly.

[auke]

Hmm that is odd, in P0 the memory should be at 7000Mhz. I would also doubt setting P0 and P2 memory clocks to 4000 will give me back the lost 40% of performance. I suspect the problem then lies in the presence of a K4000 in the system as well or the reduced PCIe lanes per card (although I doubt this as going from x16 to x8 doesn't even make a noticable difference in most games.) Could you check the status of your PCIe interface? NVinspector tells me it is at "3.0 @ 2.0 x16. Thanks

[auke]

How do I find out what P-state?

Not sure if there are any official utilities that check this but I'm using NVinspector.

[super gnu]

7000 mhz is the DDR memory rate. the actual clock to the chips is shown in nvinspector, which is 3500 mhz ( well actually i believe the actual clock is much lower than that, there are several multipliers involved. ) p2 slows the clocks to 3000 mhz, or 6000 mhz effective. so an oc of the memory to 4000 mhz gives you a final memory clock of 8 ghz

from my testing, the memory clock gives about the same, if not slightly more, benefit than a core clock increase, which is why a p0 state for RT would be quite desireable.

have you checked if your chip is throttling when you run RT? what are your temps/ clock speeds for the chip shown as in nvinspector or gpu-z when rendering?

pcie bandwith has been discussed here several times before. since rt loads the whole scene into ram, the only major penalty for a slower pcie link is longer load times, and less interactivity in activeshade, although i believe once you get down to pcie 1x levels things do start to suffer. no issue with a pcie3.0 8x link

i suspect your issue is either, as you mentioned, some driver issue related tot he quadro and the titan in the same setup, or clock throttling due to over power/overheat of the titan (although in my testing power draw when using RT is much lower than a heavy graphics benchmark such as furmark) could also be a particular issue with the driver revision you are using.


btw thanks for asking me to check my pcie link status.. for some reason my old gtx670 is running 3.0 @2.0 x16, but my titan x is running 3.0@1.1 x8 ! thats terrible. i need to check the slot spec i have it in. anyway its a good illustration as to how pcie bandwith isnt a major limit.. i get 7 minutes something on benchmark 2 on my oc titan x.