V-Ray RT 3.6

[mohammadreza_mohseni]

I did some test with V-Ray RT 3.6 and I am really satisfied with the speed and result. It was a real pleasure to work with it and we will definitely use it for our archviz stills and animations using GPU (we will invest more on GPU from now on). I most confess that I did some serious concern about stability and speed improvement of hybrid c++ CPU implantation in CUDA engine however it proved me that is a very very nice and solid implantation Chaos Group really must congratulate and thank you for this.
But, I think it needs improvements. In my tests in my benchmark scene and in some real-world, scene CPUs come to aid of GPUs when rendering time increase. In the scene, with low render time, it may not affect render time and even increase it a little bit (of course, it may depend on the scene and I only ran a few tests). Maybe, it was related to initiation time for the CPU implantation to start working (I am not sure about it).

I have capture one of my test to demonstrate it a little better. you can see the below video.
I used a machine with Intel CPU 5960X (OC, 4.1) and two GPU (one 1080ti and one 1080)
vray version was: 3.60.01

[ran_sarfati]

wow so cool hope it will be out this week so we can use it too

[savage309]

For some scenes it can be the case that having devices that are very different in render speed (like 1080ti and the CPU), because of load balancing the render to not get faster. The same will happen with the same scene if you for example use GTX1050 and GTX1080ti. Also keep in mind that the CPU has to feed the GPUs with data while rendering (the more GPUs you have, the bussy-er it is with them instead of with rendering).
The load balancing is tricky and might take some time for us to tune it well. Most of the setups that we tested here showed a decent speedup https://labs.chaosgroup.com/index.ph...rid-rendering/.
Also, the C++/CPU Hybrid render was designed for few other cases as well - reusing existing hardware / CPU servers, using V-Ray GPU CUDA on machines that do not have CUDA GPU (MacBooks while on the go, for example) and as insurance in case of very memory hungry scenes.

If you can send the scene over for investagation it will help

Best,
Blago.

[Alex_M]

Interesting comparison. I can't wait to try out 3.6. By the way, I noticed some image artifacts when using the hybrid renderer in one of the nightlies from a month ago. I hope the issue is fixed in the official release. I'm talking about thin black lines at the edges of objects.

[savage309]

It would have been nice to have the scene here, so we could have checked before the release if the issue is fixed
There was a bug with DoF & hybrid few months ago, if that was it - it is fixed.

Best,
Blago.

[Alex_M]

No, it wasn't the DOF problem. Sorry, I don't have that nightly installed anymore so I can't reproduce and send a scene. I'll check with 3.6 tomorrow and report back.

UPDATE: It seems that the issue is fixed! I no longer see those thin black lines I did in the nightlies.

[Alex_M]

I did various tests with the hybrid CUDA rendering and I saw some pretty bizarre things.

Here are my findings from the tests I did:

1. Render times didn't improve in any of the tests I conducted. In all of them the render time was either the same or longer. In one of the scenes (an interior) the render was 1.7 times longer.
2. I get noisier images when I add the CPUs to the mix and I didn't change anything in the render settings.
3. The more CPU render nodes I add to the rendering, the slower the rendering is.

I think there's something wrong with my setup or I missed some steps in order to make hybrid rendering work correctly?
Here are some screenshots:

[savage309]

Send the scene to blagovest.taskov@chaosgroup.com and peter.matanov@chaosgroup.com with link to this thread so we can check what is going on please.

Thanks,
Best,
Blago.

[Alan Iglesias]

Just some basic early observations comparing GPU/CPU/Hybrid Rendering done this morning...

On files that have always been great for GPU rendering (exterior sun/sky shots; product visualization with relatively simple "studio" lighting) I find that the Hybrid renderer can approach a 90% speed-up over using just the GPU. Pretty dang cool!

On more complex interior lighting shots I find Hybrid (using RT/GPU as the Production Renderer to make use of the Light Cache) to be about equal to my Vray Adv CPU rendering. This is still quite a speed-up as the lone GPUs rendering interior shots were always significantly slower at cleaning up noise than Vray Adv, even with the Denoiser

Note that these tests were made with rather old (but fairly powerful) GPUs and adding newer, more powerful GPUs could tip the scales significantly!

So these tests are more about relative differences in a particular system than anything else - but hopefully somewhat useful here.

As usual, YMMV...

My thanks to everyone at Chaos Group who helped to make Vray Hybrid rendering a reality...

-Alan

[Alex_M]

Send the scene to blagovest.taskov@chaosgroup.com and peter.matanov@chaosgroup.com with link to this thread so we can check what is going on please.

Thanks,
Best,
Blago.

I've just sent an email with description, scenes and a link to this thread. Let me know if you need any additional info.

[savage309]

I've just sent an email with description, scenes and a link to this thread. Let me know if you need any additional info.

Thanks!

Best,
Blago.

[super gnu]

im just doing my first tests with RT 3.6..



titan X - 1m 44

3930k@ 4.2 - 5m 24

gtx670 -3m 53

------

titan X + gtx 670 - 1m 39

titan X + 3930k - 1m 32

titan X +3930k + gtx670 - 1 m 36


does that sound reasonable? im not getting much of a speedup over a single titan x, despite the 670 and the cpu being very roughly half a titan x each. in fact all 3 together is slower than the titanx and cpu alone.


i noticed this situation was much more severe on a smaller region render, where the single titan x was significantly faster than any combination of devices.

for comparison my 3930k in ADV with the same noise thresh: 6m 45

[spohn_ranch]

I did some did some DR rendering today with 3.6 hybrid - (2 x) xeon e5-2699 v4 +(1x) Titan Xp and (4x) Titan XM in total. It is insanely fast.., once it gets going. I'm pleased with the way Vray is developing.

[M.Max]

I've tested it today and get quite good results suing :
GPU :1080Ti
CPU : 2* E5 2696 V4

Here is the results:
RT GPU 5:37 min
RT CPU 7:24 min
RT GPU + CPU 2:19 min

just to compare it with what I usually use for production
Adv CPU : 4:55 min
(there is a little deference, but overall it give a good idea about the speed )
I don;t know how you make it but it is amazing !!

[bardo]

Just a question about VrayRT.
It seems really interesting the speed boost using hybrid mode...but considering that is the RT version, the question is: are there unsupported VrayADV features yet?