Vray RT 3.30.01 Issues

[Donfarese]

I'm having some major issues with Vray RT in the new 3.30.01 Build. usage and speed issues. I have a 3930k OC with 32gig ram, AMD W7000 for display, and Titan X for Compute, running Windows 10. All we use here now is RT, so finding a fix would be great!

1) The new integrated RT running in max process doesn't work with having a secondary compute card for us, will not find the Cuda Device. I tried Changing the GPU to TCC, that had major issues, one it renders very slow, two it does away with you cards auto fan control for heat, and the card heated up and was pumping HOT air out the back of the computer.

2) Tried running it as a DR service. Issue with that is that I have to run (C:\Program Files\Chaos Group\V-Ray\RT for 3ds Max 2016 for x64\bin\vray -server) from CMD prompt every time I want to render or keep the last one open. If not it doesn't find the Cuda device also. And getting this way to work the render times are much longer than from 3.20.03. Vray RT 3.2 - 21min55sec vs Vray RT 3.3 - 55min44sec. Also this mode doesn't seem to cache textures which really sucks.

3) The new Adaptive method doesn't seem to be working properly in RT, seems much faster and correct in ADV. Also RT doesn't render the same as ADV. Both are using BF,LC and progressive for ADV. I attached the rendering below to show difference in speed and render. Notice the couch and chairs especially are much whiter in the ADV rendering.

[savage309]

Thanks a lot for the feedback! On the topics:
We will try our best to manage to find a way to get the nVidia GPUs from the 3DS Max process when you are on AMD GPU for monitor and when you are using WDDM driver, instead of TCC one. Usually one of the purposes of the TCC driver mode is to boost the performance (not the opposite). So far I have not experienced any slowdowns or overheats when using that and I use it as my main workflow. If you have any of those they might not be related directly to V-Ray RT.

I understand that running "vray -server" is annoying and as I mentioned above, we will try to workaround that issue (yet I am not completely sure that there is something we can do, beside adding an option to run V-Ray as in 3.20 and prior - meaning out of process, with all the drawbacks that approach has). Out of process for example does not caches textures by default (as you noticed) and this is one of reasons to introduce the in-process render.

I am still running some renders to compare the render times and I have a few questions about how your measurements were made.

First, thanks a lot for the scene, such scenes helps us a lot! But I think there is some kind of an issue with it at our side, since the renders I get are different from the ones you posted. Not sure if I am missing a plugin or something, so if you can help with that I would be very helpful. With this being said, here is what I got so far- first, 3.20 vs 3.30, same GPU setup, driver, same render time.

With the noise approximation statistics that I have, the noise for the 3.20 is 0.030 and the noise for 3.30 is 0.032. This makes around 6% in flavour for 3.20. The images however are different, and some of the differences are because of 3.30 rendering better. However, you are right that some of the differences, like the couch, seems like a bug. I will put that into the bug-tracking system and we will fix it.

Also, I tried RT CPU (for 43mins) vs RT GPU (for 8mins) (same settings, just changed from CUDA to CPU), V-Ray 3.30.01, i7 3820 vs Titan X, and here are the results I got.

For some reason the uploaded images resolution seems different now, but they in fact they were rendered exactly the same ...

I guess what is happening is that you have set some amount of noise threshold and max sample level and for some reason the new image sampler in 3.30.01 shots many more rays that the one in 3.20.03. I am guessing that it shot many more samples even after the noise levels from 3.20.03 were reached. So you wait longer, when the render could have in fact completed earlier.

I will talk with the guys that are more familiar with the image sampler and I will update you when I have a higher degree of confidence what exactly is happening. Probably some of its settings (noise, time, etc) now are interpret differently or something else.
In the meanwhile we will investigate the coach shader and the in-process-amd-nvidia thing.

[Donfarese]

This is so great, thank you for looking into this.

The scene has no Plugins, so there shouldn't be an issue on your end. Yes the 3.3 is having some sort of bug not rendering the couch & chairs correctly. As far as the speed, like you said, if you just let both those 3.2 and 3.3 render till the end, 3.3 will take twice as long. I'ts like the settings don't work correctly with the new sampling, and this is no good for our animations that need to render out full frames. I was expecting to render this in less than half the time looking at the comparisons everyone are doing with the new 3.3 rendering speed. Also doesn't Iray support different video cards for it's Gpu renderer, maybe you can look how they did it? I noticed procedural noise doesn't work in bump yet, I thought that was already done for this new SP?

As far as the TCC, check your heat on the GPU, for some reason the GPU disregards the auto fan for temperature, for me the Titan X card started to boil. Even if it runs right, the card will be at a steady 83 degrees which is to hot for constant rendering, which is why using a custom fan curve is smart, we run our Titan X at 64 degrees always, but this two is not capable with TCC mode.

Once again thank you for the Help!

[Donfarese]

I did a test to just see, I tried the same scene in Corona. Corona doesn't have adaptive sampling as I'm aware, but it is rendering faster than Vray 3.3 even with its adaptive sampling and even faster than 3.3 RT which doesn't make any sense. Whats going on?

[savage309]

Thanks for the intensive tests you are making. I am sure testing the beta will help us all a lot for getting better SP.

I would have preferred to wait until I have more exact answers to your questions, but they are getting too many and I may start to forget . So, in reverse order ...

I am not expert in sampling techniques (adaptive vs fixed, etc), neither in the V-Ray Production render. It will be better to post questions related to that in the Production forum, where you will get more accurate answers.
I don't know about Corona neither, but I am pretty sure that they are using advanced image sampler techniques, just like we do. Apart than that, the render results here are different. As a general rule, if you compare different render results you can get anything (in your case, you can notice for example that one of the renders is darker than the other). By the way I think in the scene setup that you have send, it is the old image sampler that is used (not the new one). I still haven't had the chance to change it and run all the renders again to compare the results ...

Also, the scene setup that you have does render differently (I checked that again) on our computers. Could you please confirm that what you have send us is expected to render fine out of the box ? We ran the scene a couple of time, we got longer renders from 3.30 and part of this was the image sampler aiming (and producing) clearer image than before. But not as long as yours. The guys checked that new behavio is the more correct one. However for the official SP we will try to change the sampler behavior for RT to more closely match the old one for particular cases like yours, so it will render shorter. However, the times you got are still not reproducible here, I guess because of the differences in the render, but I want to confirm that, so I can be sure ...

For the GPU support - we do support different video cards for GPU rendering. With 3.30.01 we added better support for AMD GCN 1.0, 1.1, 1.2 (driver 15+ required) and nVidia Fermi, Kepler and Maxwell (both OpenCL and CUDA). Also we support OpenCL running on AMD and Intel CPUs. IRay runs on nVidia GPUs using CUDA and/or CPUs only. The problem that you have with AMD GPU for monitor + WDDM nVidia GPU for rendering and in-process-render is not related to what GPUs we support for RT rendering. It is that just CUDA initialization fails, when it is invoked with this setup from Max (this is Max related, not V-Ray). Probably there is something that could be done, probably not. We will do whatever necessary to find if there is a possible workaround that we can implement, and if there is we will implement it.

Procedural bump for RT GPU however would not be added for 3.30.01, since it is not ready yet. Rounded edges on the other side most likely be there.

[vlado]

I did a test to just see, I tried the same scene in Corona. Corona doesn't have adaptive sampling as I'm aware, but it is rendering faster than Vray 3.3 even with its adaptive sampling and even faster than 3.3 RT which doesn't make any sense. Whats going on?

The Corona image is somewhat darker. If you make the lighting from the window darker you will get very similar result from V-Ray too. On the whole, we have found that for many scenes, Corona doesn't have substantial advantage over V-Ray 3.3, neither in render time, nor in scene setup time. Sometimes people also forget that V-Ray is a lot more complex renderer so I personally find it amazing that it can get even that close.

Best regards,
Vlado

[Donfarese]

I totally agree that vray is much more complex, Love vray. but I even lightened the scene brighter than the vray one and Corona rendered faster. This is not a corona vs vray competition, but for some reason with vrays adaptive sampling and corona not running adaptive sampling ( which they are talking about having for the next version) it seems weird that vray is lagging behind in render speed. Now for RT, vray was always faster! the main reason I Love vray! But I am getting really bad times on 3.3 RT, the scene I sent is taking longer than CPU, and I've even had ours Cleaned and reinstalled our workstation, still the same results. Something is not right, 3.2 was faster, and with this amazing 3.3 adaptive sampling it should be way faster for the cleanup of noise in this scene with only one light. Is there anymore tests I can do that would help? I've even removed all textures, but still a lot slower than what was expected or last version. Could it be it's multiplying some of the settings more than the last version (more rays, more samples, for same settings)? Something really seems off while working with it.

[vlado]

I totally agree that vray is much more complex, Love vray. but I even lightened the scene brighter than the vray one and Corona rendered faster. This is not a corona vs vray competition, but for some reason with vrays adaptive sampling and corona not running adaptive sampling ( which they are talking about having for the next version) it seems weird that vray is lagging behind in render speed.

<shrug> Then obviously we have more work to do

But I am getting really bad times on 3.3 RT, the scene I sent is taking longer than CPU, and I've even had ours Cleaned and reinstalled our workstation, still the same results. Something is not right, 3.2 was faster, and with this amazing 3.3 adaptive sampling it should be way faster for the cleanup of noise in this scene with only one light. Is there anymore tests I can do that would help?

Just please help Blago to get the scene rendered in the same way as it renders for you, and we'll take it from there. Thanks for helping on this!

Best regards,
Vlado

[Donfarese]

Will run some more tests, and provide more scenes if needed. Love what you are doing with vray.

[Donfarese]

when I render with CPU, what is this Prepass it is doing? I thought that wasn't activated on this build because it wasn't finished?

[vlado]

when I render with CPU, what is this Prepass it is doing? I thought that wasn't activated on this build because it wasn't finished?

What does that prepass look like?

Best regards,
Vlado

[Donfarese]

Looks like an IR pass but dark.

[Donfarese]

Did another test where I rendered with half the Trace depth, and Half the GI depth too see how much faster it renders. It actually rendered slower.....? trying to figure out what is going on... and it is still rendering over twice as slow as 3.2. Check images attached.

[savage309]

I am afraid there are a some stuff which I can't answer at the moment, since I don't understand well what is going on ..
Even if we leave for the moment the issues we have to render your scene (I showed earlier how it is rendered at our side).
This one for example http://forums.chaosgroup.com/attachm...9&d=1447373211 should show 3.20 GPU vs 3.30 GPU. In 3.30 the bump of the floor should be different, as well as the couch (it is rendered white in 3.20 and yellowish in 3.30 in the scene you have sent) and some of the refractions/reflections (because of bugs being fixed in 3.30). The renders you have shown however are exactly the same. Probably you put the same image twice by mistake ? Or this is correct and this is how it renders at your machine ?
In the images you have shown you are also comparing 20 minutes CPU render vs 40k samples GPU render. The CPU render clearly is noisier and to get a proper result, you should set the both engines to equal conditions (otherwise you can just set 10 minutes render time limit to the GPU render and it will be 2 times faster than the CPU, for example).
And I am not quite sure what that prepass is. The scene as you have send us is set with Light Cache. Could this be the LC prepass (I am not sure it looks like LC, but it will be good to check that) ? Can you try switching the GI engines to BF/BF to see if the prepass will still be there ? Also, your scene was using the old adaptive sampling algorithm, not the new one ...
For the num bounces - V-Ray RT will use the minimum of the values in the RT settings and in the materials themselves (so in your case if your materials are having limits <10, it would not matter). And are you sure these are the only changes in the settings that had been made ?
But in the end, I am just guessing (most likely, wrong-guessing), since the scene you have send renders differently for us.

If it is possible, I would like to ask you to pack and send us the scene again, just in case.
I understand that this might be time consuming for you, but if we look at the same thing it could make it easier.

[Donfarese]

Okay thank for the feedback, I know you can't answer every question right now, I wouldn't expect you to since you just released this beta. Just trying to get you examples and info back for help for you guys to figure it out. I also forgot to state that I also dropped the max pixels/paths from 40000 to 10000 on the render that should of been faster but was 3min slower, weird! I will have two new scenes for you to use, same scene one RT one ADV. In the ADV I am just using defaults, except I changed the render time to 0, and reduced Light Cache Subdiv to 300 to match the RT and turned off LC Leak prevention. LC Leak prevention always did the opposite for me in RT, it would leak light in edges and corners unless I turned it off. I just emailed both scenes to blagovest.taskov@chaosgroup.com email, there in a Dropbox.

Here is a shot of ADV with BF,BF. same prepass. ?