KeyShot Forum

Technical discussions => General discussion => Topic started by: Jet3D on May 05, 2020, 10:17:12 AM

Title: SLI GPU vs 3950X render results - maybe a point of interest
Post by: Jet3D on May 05, 2020, 10:17:12 AM
Hi guys and ladies,

I thought I would share my results for testing KS 9.3 on a test bed system.
I am conducting tests for work during lockdown...

This may all be common knowledge, (I don't visit much) however, I thought that I would share as it may help others.
Amazing work, Keyshot team. GPU rendering crushes the CPU results -  I can't wait to see two or more GPU's harnessed for speed.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DerekCicero on May 05, 2020, 12:06:13 PM
Good stuff. Do you have NVLink (it seems to show it on the right)? KS 9.2 and above support NVLink.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 06, 2020, 12:20:50 AM
For this scene in KeyShot 9.3, with NVLink, you should be seeing a nearly doubling of speed when using two cards instead of one.

Are you sure SLI is configure correctly in NVIDIA Control Panel ('Maximise 3D performance')?
Are you sure both GPUs are selected in KeyShot ('GPU Usage')?

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: Jet3D on May 06, 2020, 11:29:20 AM
Hello Derek,

To answer your question - yes, the cards are NVlinked via an ASUS bridge.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: Jet3D on May 06, 2020, 11:33:00 AM
DriesV,

Yes in the Nvidia panel - it is set to maximum with both cards in SLI.
I tested many time with this enabled and disabled.

I will test again but, I am certain GPU usage was selected. I will return and let you know in an hour or so.  8 )

Regards, Jet
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: Jet3D on May 06, 2020, 12:03:22 PM
Update with a five second speed difference. Same settings - not sure what changed unless it was a background task.....
Not a great increase - your thoughts?
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 07, 2020, 05:48:43 AM
Maybe the rendering is just too fast? :)
What results do you get with, say, 1024 instead of 256 samples?

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: mafrieger on May 08, 2020, 05:02:49 AM
Quote from: DriesV on May 06, 2020, 12:20:50 AM
For this scene in KeyShot 9.3, with NVLink, you should be seeing a nearly doubling of speed when using two cards instead of one.

Dries

hmm just a question:
whats the technical background for doubling speed in this case?

when asking whats for "what's NVlink good for?", the answer
was "only doubling memory"
QuoteNVLink will not double speed though. It is not a magical solution with linear performance scaling.
please see https://forum.keyshot.com/index.php?topic=25847.msg108674#msg108674

were there big changes regarding this when moving from 9.2 to 9.3?



Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DerekCicero on May 08, 2020, 10:01:47 AM
Dries can provide additional detail on Monday, but I believe the answer is that is when the memory stays under the individual card threshold, performance will double, however once you pass the threshold on one card (say 8GB), normally it would throw up an "Not enough memory" warning , but with NVLink, it will kick over and pool the memory on the cards (now 16GB) so you can do the render, but at that point you lose the doubling effect b/c the cards have to switch to the process the memory sharing, swapping all the textures, etc.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: Jet3D on May 09, 2020, 04:40:31 AM
These are 11GB cards so, it would have a total of 22GB.
We have some Quadro P6000 24 GB cards at work. I will test with these when lockdown lifts.

I have been CPU rendering and if I had only tried GPU at some stage - 95% of the time, I would never go back.

After what I have seen that KS 9.3 can do - I will be steering work into using KS in the future. Perfect for what I do and able to deliver fast results in the film industry that needs everything yesterday! LOL.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 10, 2020, 02:12:54 AM
It seems like you are able to render the scene on a single card. In that case, performance should nearly double with two cards.
What I said in that thread was just a heads-up that you shouldn't expect double performance in all cases.

You don't happen to be able to share the scene, do you? Then we could check internally.

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 10, 2020, 02:16:40 AM
In hindsight, are you actually sure you are just rendering with a single card? Turning off NVLink doesn't mean that you can't use both cards in KeyShot. It just means that the memory won't be pooled, as Derek outlined.
Enabling NVLink will not do anything for perfomance. It just allows you to fit larger data sets in both cards.

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: mafrieger on May 10, 2020, 05:00:47 AM
Quote from: DriesV on May 10, 2020, 02:12:54 AM

What I said in that thread was just a heads-up that you shouldn't expect double performance in all cases.


Many thanks for clarifying this !!

Some questions regarding usage of GPU's VRAM:
1) what's biggest VRAM amount needed for a scene you have seen so far?
2) What are the things that drive VRAM amount up most?
3) Could you share a demo scene with very big VRAM needed?
With this one (jet3d an others) could see if a / a couple of Quadros with losts of VRAM is worth it / needed (in real live)...
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: menizzi on May 10, 2020, 09:14:16 PM
I must be doing something wrong. My gpu renders are really really really bad. 100%  unusable bad. as in so bad I bought a lic for octane standalone and i am trying to learn that but i have not had much time.

Maybe in Version 10? If you click the Denoise button watch out its so bad it makes me cringe. I also have a 2080 super.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 11, 2020, 03:50:50 AM
Quote from: menizzi on May 10, 2020, 09:14:16 PM
I must be doing something wrong. My gpu renders are really really really bad. 100%  unusable bad. as in so bad I bought a lic for octane standalone and i am trying to learn that but i have not had much time.

Maybe in Version 10? If you click the Denoise button watch out its so bad it makes me cringe. I also have a 2080 super.

Let's try to break your issues down.

What kind of scene are you trying to render?
Is it Denoise specifically that is bad?
Can you pinpoint the issues you are running into?
Can you share your problematic scenes perhaps?
What is it you think Octane will/might do better than KeyShot?

Let's try to be constructive and find a solution. :)

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: menizzi on May 11, 2020, 08:39:15 AM
I can share the scene with you not the public. I have a deal that i will not share it online. I think the guy might be dead i am unsure as i ruined the relationship we had at the end. he got covid-19 and i was just worried about getting the work i paid for. Having said that.

The file is 6.3gb in size. Were do you want me to upload it for you? I am rendering it out now via GPU for you but I am going to let it go to stupid samples. My biggest problems

1. Can not get cpu quality or even close to it via gpu
2. Substance source materials can not be used.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: menizzi on May 11, 2020, 06:17:13 PM
Gpu on a windows 2080

(Edit) I see it did not screen shot the display screen. That is the best it could do with 75,000+gpu samples and about 3 1/2 hour render time. Like i said I have no idea how you guys are getting the gpu to give the same results as the cpu.
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: DriesV on May 11, 2020, 11:58:03 PM
Without seeing the CPU image, it is hard to tell what is wrong with this GPU rendering.

For sharing the scene, so that we can have a look:
You can send data securely via our WeTransfer (https://keyshot.wetransfer.com/). Send to dries-at-luxion-dot-com.

Dries
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: mafrieger on May 18, 2020, 12:54:43 AM
Quote from: mafrieger on May 10, 2020, 05:00:47 AM
Quote from: DriesV on May 10, 2020, 02:12:54 AM

What I said in that thread was just a heads-up that you shouldn't expect double performance in all cases.


Many thanks for clarifying this !!

Some questions regarding usage of GPU's VRAM:
1) what's biggest VRAM amount needed for a scene you have seen so far?
2) What are the things that drive VRAM amount up most?
3) Could you share a demo scene with very big VRAM needed?
With this one (jet3d an others) could see if a / a couple of Quadros with losts of VRAM is worth it / needed (in real live)...

=> just put this into a new topic: https://forum.keyshot.com/index.php?topic=26232.0
Title: Re: SLI GPU vs 3950X render results - maybe a point of interest
Post by: smalldogstudio on May 18, 2020, 04:24:28 AM
Quote from: Jet3D on May 09, 2020, 04:40:31 AM
These are 11GB cards so, it would have a total of 22GB.
We have some Quadro P6000 24 GB cards at work. I will test with these when lockdown lifts.

I have been CPU rendering and if I had only tried GPU at some stage - 95% of the time, I would never go back.

After what I have seen that KS 9.3 can do - I will be steering work into using KS in the future. Perfect for what I do and able to deliver fast results in the film industry that needs everything yesterday! LOL.

Thanks for posting the test - interesting! - my own experience is my relatively humble RTX2060 was 16.8x my CPU for  rendering. Considering an upgrade of either an identical 2060 or something better

bw

David