Lightcache taking forever with vraydisplacementmod 3d

[daltonwatts]

Hi guys!

 

I'm having a problem with a small scene file on my new 2x Xeon 2696v3 workstation. I didn't had the same problem on my old single Xeon 1230 v2 Workstation.

 

On the old workstation i was running 3ds max 2014 with Vray 3.2 and on the new workstation i upgraded to 3ds Max 2016 and Vray 3.3.

 

I already applied the VRAY_USE_THREAD_AFFINITY=1 environment variable for vray to be able to see more than 1 cpu while rendering, so no problems there.

 

Problem is: If i turn on vraydisplacementmod 3d mapping type the lightcache takes forever to be calculated. If i change it to 2d type it renders fast, no problems at all. I'm on Windows 7 Ultimate x64.

 

EDIT: If i turn OFF the VRAY_USE_THREAD_AFFINITY=1 environment variable so as to use only one cpu, vray resolves light cache faster than if using the 2 cpu's... Like 3x to 4x times faster but still in an unnaceptable speed. This is REALLY weird!

Edited June 19, 2016 by daltonwatts

[Francisco Penaloza]

LC is not displacement friendly, if there is many object with it or very large areas I would recommend to use Brute force.

[jasminepeterson]

Displacement mapping is a technique for adding detail to your scene geometry without having to model it first. The concept is very similar to bump mapping. However, bump mapping is a shading effect that only changes the appearance of a surface, while displacement mapping actually modifies the surface.

[daltonwatts]

I know what displacement and bump mapping mean. This is a weird issue because i don't have the same problem on a much weaker workstation. It renders very fast even with 3d type enabled. I guess i'll have to go back to max 2014 and vray 3.2 because this doesn't make any sense whatsoever.

Edited June 20, 2016 by daltonwatts

[Francisco Penaloza]

Not sure what your problem is really but some think to consider.

 

Your old CPU was actually faster 3.3 Ghz, this will really help in single core process such, view port, scene translation and so on.

Your new Computer setup 2x Xeon 2696v3 is only running at 2.6Ghz, this should render images faster just because it has more cores. But in single core operations will be slower.

 

To my understanding the variable (VRAY_USE_THREAD_AFFINITY) it is useful when you have more than 64 cores in a single machine.

in your case you max is 56 so you are still under default values.

 

3D displacement uses more memory and CPU than 2D displacement. Light Cache is slower resolving large or complex displacement. Both LC and Displ are memory chugger, check your memory while rendering, maybe you are running out of RAM and putting all that stress in the scratch disk, which will slow thing down to a crawl.

[daltonwatts]

I'm not running out of RAM that's for sure. Light cache is faster on the new workstation than on the older one with 2d displacement type enabled. I'll really have to reinstall Max 2014 and vray 3.2 on the new machine and see how it goes because this makes no sense

Edited June 20, 2016 by daltonwatts

[daltonwatts]

Ok, if i'm using the 2 cpu's with the environment variable enabled and tick the "cache normals" and "static geometry" on vraydisplacementmod, it takes 2min to compile geometry and 20seconds to render the entire scene.

 

If i remove the environment variable and only use one cpu it only takes 20seconds at compiling geometry and renders the scene in 35seconds. This is using Irradiance map. My old single Xeon renders the scene in 4m55.

 

I seriously don't get this. Does vray do a bad job in managing 72 threads? Corona has no trouble in utilizing all cpu's and renders the same scene in 1min.

 

I must say that i also tried vray versions 3.40.01 and 3.40.02. 3.40.02 crashes max before starting the render. 3.40.01 has the same problem of 3.30.05.

Edited June 20, 2016 by daltonwatts

[daltonwatts]

Ok, some more tests done:

 

On i7 4770k | 8gb RAM (onboard VGA)

Max 2014 vray 3.30.05

Opens the scene in 24 seconds

 

On Dual Xeon 2696 v3 | 32gb RAM (Quadro K1200)

Max 2016 vray 3.30.05

Opens the scene in 54 seconds

 

After this i thought i should revert back to Max 2014 at least to see if opening times would decrease. Results:

 

On Dual Xeon 2696 v3 | 32gb RAM (Quadro K1200)

Max 2016 vray 3.30.05

Opens the scene in 45 seconds

 

Not much of an improvement but i guess cpu frequency might play a factor here?

 

Now even more strange:

 

i7 4770k with vraydisplacementmod 3d type enabled resolves lightcache in 17 seconds and renders the scene in 2minutes and 20seconds.

With Dual Xeon 2696 v3 with 3d type enabled the lightcache takes 4minutes and 10 seconds to resolve...!!!!!!!!

 

This is with cache normals and static geometery turned OFF.

 

With 2d displacement type i have no problems in resolving light cache. i7 4770k renders the scene in 2:09 and Dual Xeon renders it in 26seconds.

 

One of the Xeon cpu's is an "ES" version. Does it play a factor here when using 3d type?

Edited June 21, 2016 by daltonwatts

[Vince Paske]

Turn your dynamic memory limit up to 4,000 (found in the VRay settings tab).

[daltonwatts]

Already done that Vince. Turned to 12000, 10000, 8000, same result.

[Francisco Penaloza]

CPU Frequency plays an important factor here, also with high number of cores there is a very small delay in translation, most of the time is almost none existent but depending of the scene it may be a little more noticeable.

From all your test, you can tell that the high frequency CPU will translate and process faster, now when actual rendering kick in, then when more cores you have the better.

If you want more detail about the tech behind displacement and Light cache, I would recommend to post this at Chaos group web forum. There Vlado or some other tech savvy member can give you more inside about it.

 

Regarding your comment comparing Corona to V Ray, yes Corona Tech is 'newer" than V-Ray. Chaos group in the latest release adjust the code to fix all those issues with large amount of cores in a single machine.

 

if you run Cinebench test in both machine, you'll get a good comparison of speed of single cores processing and multi core processing.

 

P.S.

be aware that there will other factor that may affect your test, such only 8Gb RAM in your i7 machine compared to 32Gb on your Xeon machine.

Displacement and Light cache are RAM chuggers.

[daltonwatts]

Hi Francisco, thank you for replying. RAM isn't an issue in this scene. It's a small scene (60mb). The 8gb RAM machine translates quicker than the 32gb machine...

 

Vray 3.40.02 DEMO and 3.40.01 DEMO both have the exact same issues as 3.30.05.

 

Cinebench was the first thing i tried with the new workstation. 4398CB. This i7 4770k if i remember correctly gets 620CB.

 

Nothing of this matter happens in Corona. In Corona the i7 4770k renders the scene in 6:31 and the Dual Xeons render in 1:10. No delay in translating or compiling like vray...

Edited June 21, 2016 by daltonwatts

[Francisco Penaloza]

I can not give you my opinion regarding the Demo version. They are limited in some functions, but again I don't know how much.

 

Cinebench can test multi-core and single core CPU, those values are for multi-core CPU. run the full test and you'll see the difference.

 

Corona is a different tech, it is also using progressive rendering. you can try That on V-Ray too, but again, their way to calculate GI and raytracing is different.

 

I used to have a work station with 2 x Xeon E5645 @ 2.6 Ghz and now we switch to single CPU Xeon E5-1650 v3 @ 3.5 Ghz

Over all this new machine 'feels' faster, opening app, quick small renderings, over all viewport performance.

But still the old workstation render faster, because it has more cores. We use it as render node now.

[daltonwatts]

Yes, single cpu benchmark in cinebench was something i haven't tried until now:

 

i7 4770k Single - 154

Xeon 2696 v3 Single - 145

 

It's not a big difference in single thread so this must not be the main cause. It doesn't explain the massive difference in calculating light cache between the 2 cpu's. 17 seconds for i7 4770k and 4 min and 10 sec for the Xeon's. Not to mention the actual bucket rendering was buggy on the Xeon's.

[Guest dialog]

Upgrade to 3.40.02

[daltonwatts]

Vray 3.40.02 DEMO has the exact same issues.

[Guest dialog]

Can you post some screen shots of your render settings. If you zip a file up and send to me, I have a few minutes I can take a look. Use wetransfer.

[daltonwatts]

Thank you Desmond but i don't think it's a problem with the file or render settings. It's a specific problem with 3d displacement and my Xeon's. It renders just fine on my i7 4770k. If the same file works on a much weaker cpu it should also work on the dual xeon's. I tried everything by now.

 

I'll have to try other older scenes with 3d displacement to see if the problem is the same.

[Guest dialog]

Sounds good.

[daltonwatts]

Well, it seems i found out the problem.

 

Changing edge length from 2.0 to 5.0 and ticking ON both the "use object mtl" and "view dependent" makes light cache resolve in 5 seconds I can also untick both "use object mtl" and "view dependent" with 5.0 in edge length but it will render about 2 times slower. The problem is if i untick "use object mtl" and keep "view dependent" ticked with 2.0. That makes light cache veeeeeery slooooooow. This is all with cache normals and static geometry unticked.

 

Thank you for all of you that took the trouble to reply