Soba Noodles Benchmark scene (CPU and GPU)

[Ivan1982]

Hello,

I don't know if i am at right sub-section of the forum regarding this topic, i apologize if i am not, please do what you have to, thanks.

I have created this benchmark scene to share with all of you guys. It is a user bench scene for Vray 4.3 (max 2018 and above). It was created solely for this purpose which is to compare speeds between those 2 engines,

About scene:

This is user bench scene, basically real life scenario type of benchmark it is free for download so you can test your machine or you can use it how ever you want commercially or not.

Download link (around 2gb) https://mega.nz/file/dEUXjABb#TkiUI0...-HqKN1WpohVWD8

You can UnRar it on your C drive, there shouldn't be any issues regarding loading of profiles for lens effects and VFB settings or wherever and load manually.

There are Vray CPU (ADV) version and Vray GPU (CUDA) version, there are slight difference between engines however i made sure (as much as i can) that noise levels are comparable between engines.

2 views, one wide angle and one close up just to mix things up i mean i never made bench before so pretty much don't have any idea what i am doing

Everything is more or less at default values except reduced samples and bucket size for CPU. Scene loading can last few seconds since all textures are showed in viewport, please be patient.

This should run on GPU with minimum of 6GB Vram (GTX 1060 laptop) works without any texture resizing. I didn't test RTX since i do not have RTX card and also i didn't try it on GPU with lower amount of Vram since i do not have card with less than 6GB, if it works that's great.

Resolution is 1500pix x 1500pix since i wanted to keep results in minutes, i guess TR 3990x or 2-3 x 2080ti will finish before they start

I have tested noise levels to make sure there are more or less consistent in both scenes and make render adjustments according to them to produce similar noise level images, once again there are subtle differences of course.

Corners of images are measured so is center of the image around Irori.

Noise levels check with NEAT denoiser:



This bench also includes Vray Lens effects (NOT hardware accelerated though), Nvidia AI Denoiser and LUT corrections (created for this scene), all applied in VFB.

My home WS specs:
Intel i9 7940x
96GB of RAM
1080ti Founders Edition
All stock frequencies.

Here are my results.

VRay CPU test:



Render time View 1 CPU: 18m 25s



Render time View 2 CPU: 12m 05s

VRay GPU Test:



Render time View 1 GPU: 12m 33s



Render time View 1 GPU: 12m 42s

Images you will get are denoised, it does not have any impact on render time except of course denoising process itself, which is pretty fast anyways.

Regarding Vray 5, i will update it later or if anyone want he can do so himself, however currently there is some scene render initialization delay (it can last up to minute on my WS) with Vray 5 and results are slower so i will wait to see if that will iron out in next week or so.

Enjoy rendering.
Ivan

p.s.
https://en.wikipedia.org/wiki/Soba

[Paul Oblomov]

Until GPU DOF is fixed - I think, there's no purpose to compare You won't get 1:1 images.

[Ivan1982]

I can't say i quite understand you.

However, CPU and GPU engines are 2 different engines with in between compatibility, in either way you can achieve similar results.

[Paul Oblomov]

No you can't. GPU DOF works differently.

[seandunderdale]

....works differently, but looks different? If so, how different? Im used to GPU looking different, but not wrong.

[Ivan1982]

Paul Oblomov Now that i know what you talking about it's easier for me to respond you.

Everyone is able to use scene in way they choose no question about it, however there are lot's of people who would like to know how much faster something can be rendered if they for instance choose to upgrade their system, and what decision for upgrade to make, either CPU or GPU and what is more cost effective to them. This is main objective of this scene.
So many things differ between the engines and there are equal amount of workarounds and probably they are two different engines for a reason.
This is something chaos group state every time they can, and probably same reason they do not dare to do direct comparison, maybe because there are lot of people out there who would come to wrong conclusion and they i guess want to avoid all that calamity and responsibility.

It seems to me you didn't even try to read description which is clear (in second line original post states main purpose of this benchmark), did you download scene or at least images to see your claim is true or completely random and non related?

So here are 2 images and scene attached:





Apart from obvious difference of how light travel and bounces differently between 2 engines (CPU and GPU) which is outlined on image bellow, which is again very observable in original scene also.


You can observe that subtle or not so subtle differences by downloading both images and switch between them, which has nothing to do with depth of field you are talking about.

Is this what you stated in previous post? I cant say DOF is any different in this example but since i rendered over 100+ projects on GPU i do tend to have reserved stance, and i know that CPU and GPU have differences, some less and some more pronounced.

That is not the case here, even if there is difference, it is so minuscule that is hardly something that calls "Until GPU DOF is fixed - I think, there's no purpose to compare" statement.

Teapot scene attached so you can test it yourself : SCENE

edited to set full size images

[Paul Oblomov]

It is different on how bokeh are created. See yourself.

[Ivan1982]

No it's not. It is completely same, but you won, i am done here regarding disproving your claims.

CPU:

GPU


Purpose of my thread is to share with people something they can all use and compare benefits they can experience. There are people who are beginners reading forum, who want to learn and maybe this will help them choose how to build their system and whatnot.

[Paul Oblomov]

Heh. You showed exactly what I was talking about Hope you know, how proper DOF looks like

P.S. Scenes attached. Guess, where proper DOF is.

[M.Max]

Thanks for putting this together!
Even if it is not perfect match between these 2 engines, it is very useful to compare the "current speed" imo .
I'm downloading the file now and I will post my benchmark tomorrow.
I think making a table on google sheets will make it easier to sort all the results .

btw what's you hardware for these renders?

[seandunderdale]

Yeh, thanks for the hard work! I use Maya but its all transferable info. Itll be good to get plenty of render comparisons, even as mentioned, it might not be 100% accurate.

[Ivan1982]

You guys are welcome, i will also release Blender version (i hope so), my config Intel i9 7940x 14 core, 96gb or ram, 1080ti founders edition.

[M.Max]

So I did render 2 renders for the main View ( cam 24 mm ) , I'm not sure if that would affect the result that much but I'm using Vray 5.0 beta ( 2-5-2020 build)
my hardware is :
dual xeon 2650 v3 , 64 Ram and one 1080 ti

CPU : 36m 37 s
GPU:1h 2m 23 s

I noticed that there is a denoiser element set to AI Nivdia mode, but as I have issues with this desnoiser I changed it to VRay denosier .

I will render the other view during the night today and update the post here.

Ivan1982 How many GPU did you use to get that 12 minutes render ! ? I'm a bit jealous now =D

[Ivan1982]

That is a bit strange, I have only used one GPU (have one in WS), i did notice some strange issues with Vray 5 (i mentioned that in first post), do you want to try with Vray 4?

I am getting similar render times in Blender also.

[M.Max]

That is a bit strange, I have only used one GPU (have one in WS), i did notice some strange issues with Vray 5 (i mentioned that in first post), do you want to try with Vray 4?

aha ,that's interesting !
I will down-grade to V4.0 now to check .
the difference is huge indeed ,so there might be a bug in V 5.0 !

Update:
I installed V 4.30.02 and render time went down to 34m and 27 sec !
it is still not as fast as yours, but I'm reporting this to support team to check why there is this huge difference.
I will download the latest stable V 4.30.03 to check if that would make any difference too .