Page 1 of 2
GPU Comparison Tests
Posted: Wed Feb 14, 2018 5:07 pm
by dmeyer
I'm testing multi-GPU render in some other rendering packages as a way to understand the relative performance various tiers of cards. I am using Redshift on this test as Maxwell does not yet support multi-GPU. One thing to note about Redshift is it does not punish GPU's quite like maxwell or v-ray during renders so the real world differences between the cards will be a bit more pronounced with those.
1x 1080 (GP104x2560) 24:59
2x 1080 (GP104x2560) 12:39
1x Titan Maxwell (GM200x3072) 28:35
1x Titan Pascal Xp (GP102x3840) 19:29
2x Titan Pascal Xp (GP102x3840) 10:33
3x Titan Pascal Xp (GP102x3840)
1x Titan Pascal X (GP102x3584) 23:41
2x Titan Pascal X (GP102x3584) 12:40
3x Titan Pascal X (GP102x3584) 8:35
1x 1080Ti (GP102x3584) 21:17
2x 1080Ti (GP102x3584) 10:53
3x 1080Ti (GP102x3584) 7:38
1x Quadro P5000
2x Quadro P5000
3x Quadro P5000
4x Quadro P5000
Re: GPU Comparison Tests
Posted: Wed Feb 14, 2018 5:22 pm
by dmeyer
Random observations:
It varies but GPUs tend to max out around 65% of rated TDP for power usage during rendering.
System: 1950x CPU, 32GB Ram, 3 1080Ti: 550w peak on 110v during GPU-only render.
System: 1950x CPU, 32GB Ram, 2 1080Ti: 400w peak on 110v during GPU-only render.
Re: GPU Comparison Tests
Posted: Wed Feb 14, 2018 5:49 pm
by Mihai
Very informative, thanks again! It's encouraging to know the cards won't draw near their total power rating, at least with Redshift...and Maxwell multiGPU is coming soon...
I think, whether it's 110 or 220V shouldn't matter as far as power consumption is concerned? Same amount of energy is consumed, just that in the US you have double the amperage feeding the power supply, to make up for the 1/2 voltage. In other words a 1200W power supply whether for the US or Europe will be capable of supplying at peak, 1200W of power.
Re: GPU Comparison Tests
Posted: Wed Feb 14, 2018 5:56 pm
by Pomrose
Is it worth doing a test of the same scene with a thread ripper and a 7980XE
to see a GPU vs CPU comparison?
Re: GPU Comparison Tests
Posted: Wed Feb 14, 2018 6:19 pm
by dmeyer
Pomrose wrote: ↑Wed Feb 14, 2018 5:56 pm
Is it worth doing a test of the same scene with a thread ripper and a 7980XE
to see a GPU vs CPU comparison?
This particular scene is in Redshift which does not run on CPU.
I will try some other tests on Maxwell with single GPU vs CPU.
Re: GPU Comparison Tests
Posted: Wed Feb 14, 2018 6:43 pm
by dmeyer
Using "Bottles_On_A_Window" test scene in Maxwell 4:
Using Maxwell Render 4.2.0.2, open scene, only change to GPU engine - run to SL16
CPU tests using CPU Production Engine + Priority:Normal
Titan X Maxwell: 12:52
Titan X Pascal: 8:33
Titan Xp Pascal: 6:49
1080Ti: 7:29
Threadripper 1950x: 7:50
2x Xeon E5-2697v3: 4:54
2x Xeon E5-2699v4: 3:21
2x Xeon Gold 6154: 3:12
Re: GPU Comparison Tests
Posted: Thu Feb 15, 2018 1:58 pm
by arcmos
@dmeyer
Can you please do the bottle scene with 1 P5000?
Are you using the new Vray benchmark test?
Re: GPU Comparison Tests
Posted: Fri Feb 16, 2018 2:24 am
by Max
dmeyer wrote: ↑Wed Feb 14, 2018 6:43 pm
Using "Bottles_On_A_Window" test scene in Maxwell 4:
Using Maxwell Render 4.2.0.2, open scene, only change to GPU engine - run to SL16
CPU tests using CPU Production Engine + Priority:Normal
Titan X Maxwell: 12:52
Titan X Pascal: 8:33
Titan Xp Pascal: 6:49
1080Ti: 7:29
Threadripper 1950x: 7:50
2x Xeon E5-2697v3: 4:54
2x Xeon E5-2699v4: 3:21
2x Xeon Gold 6154: 3:12
uhm what? sorry am i reading this right? A titan X is slower than a threadripper on maxwell rendering?
Re: GPU Comparison Tests
Posted: Fri Feb 16, 2018 8:41 pm
by dmeyer
Max wrote: ↑Fri Feb 16, 2018 2:24 am
dmeyer wrote: ↑Wed Feb 14, 2018 6:43 pm
Using "Bottles_On_A_Window" test scene in Maxwell 4:
Using Maxwell Render 4.2.0.2, open scene, only change to GPU engine - run to SL16
CPU tests using CPU Production Engine + Priority:Normal
Titan X Maxwell: 12:52
Titan X Pascal: 8:33
Titan Xp Pascal: 6:49
1080Ti: 7:29
Threadripper 1950x: 7:50
2x Xeon E5-2697v3: 4:54
2x Xeon E5-2699v4: 3:21
2x Xeon Gold 6154: 3:12
uhm what? sorry am i reading this right? A titan X is slower than a threadripper on maxwell rendering?
On this scene, at this time, yes.
Re: GPU Comparison Tests
Posted: Sun Feb 18, 2018 9:31 pm
by Mihai
Max wrote: ↑Fri Feb 16, 2018 2:24 am
dmeyer wrote: ↑Wed Feb 14, 2018 6:43 pm
Using "Bottles_On_A_Window" test scene in Maxwell 4:
Using Maxwell Render 4.2.0.2, open scene, only change to GPU engine - run to SL16
CPU tests using CPU Production Engine + Priority:Normal
Titan X Maxwell: 12:52
Titan X Pascal: 8:33
Titan Xp Pascal: 6:49
1080Ti: 7:29
Threadripper 1950x: 7:50
2x Xeon E5-2697v3: 4:54
2x Xeon E5-2699v4: 3:21
2x Xeon Gold 6154: 3:12
uhm what? sorry am i reading this right? A titan X is slower than a threadripper on maxwell rendering?
From what I understood about GPU rendering in general - the more complex the light paths, the less efficient a GPU will be, compared to a CPU. A GPU has a lot of less sophisticated cores and simple computational tasks like gaussian blurring a 2D image is really fast on a GPU because you could assign each pixel to a separate core. But when you're dealing with path tracing, caustics, a lot of indirect lighting, the GPU will lose this efficiency.
But I still see substantial benefits with Maxwell on the GPU though, even if your high rez render won't even fit on a GPU, or rendering on a single GPU won't be much faster than on the latest CPUs for difficult renders:
-much faster and smoother material editing, so you can create better looking materials and finish scene tweaking faster before a final render.
- much cheaper to add 3x GPUs in your main computer and have at least 3x faster renders (and tweaking) compared to buying 3x complete AMD Threadripper computers.
- for most renders, the GPU will be faster than CPU and for exterior or product renders I think the GPU will be MUCH faster. One test I did of an exterior rendered to SL 19 on CPU (i7 6700K) in 59 minutes, on my GPU (a 1070) it took 13 minutes. A 1080Ti would have probably finished it in 5-6 minutes.
Re: GPU Comparison Tests
Posted: Tue Feb 20, 2018 11:40 pm
by dmeyer
For perspective on the GPU front...and how much it will depend on scene...
I just dusted off a production scene from a couple years ago. Product shot, emitter, displacement, etc. High res and run to SL 18
14-core i9 @ 4.4 Ghz - stopped it after a half hour but it was estimating 16 hours
single Titan Xp: 50 minutes
Re: GPU Comparison Tests
Posted: Thu Feb 22, 2018 6:02 am
by dmeyer
The unicorn:
Re: GPU Comparison Tests
Posted: Thu Feb 22, 2018 5:04 pm
by Mihai
Only 7000?
It looks like the perfect card for very high end video editing, smooth instant scrubbing of
8K video kept in the 2TB of internal memory...lets see for rendering...
Re: GPU Comparison Tests
Posted: Thu Feb 22, 2018 5:43 pm
by dmeyer
Mihai wrote: ↑Thu Feb 22, 2018 5:04 pm
Only 7000?
It looks like the perfect card for very high end video editing, smooth instant scrubbing of
8K video kept in the 2TB of internal memory...lets see for rendering...
Street price is around $4000 USD here. This would be used in 8K .r3d edit suites if performs as advertised.
Re: GPU Comparison Tests
Posted: Thu Feb 22, 2018 9:31 pm
by JDHill
I am running some blue AMD cards too...
