Differences between RT GPU and 'normal' Vray

[tricky]

I've just been messing around with RT GPU to see what it can and can't do. I need to be sure that the renders from RT GPU match (as closely as possible) the final renders using normal Vray.

The attached scene shows quite dramatic differences between the two.

I am using a vray sun/sky, some simple geometry and a vray2sidedmtl with various random settings.

test scene rt 01.zip

[tashko.zashev]

Hi Tricky,

Thanks for posting the scene file, it helps a lot to investigate this issue.

Found few issues/differences between RT GPU and production renderer, that need to be fixed or supported.
*) Color and Fade variance for "Tile" map are already added in our "to do" list
*) Unfortunately 2 sided material is not supported by RT GPU yet.
*) "Trace reflections" option for the V-Ray Mtl is now added to our bug tracking system.
*) The same thing with "Falloff" map, when is set to the V-Ray MTL Refraction slot
I hope this information helps with this scene setup.

[tricky]

Thanks for checking through this. We are hoping to put rt GPu into our production pipeline so we need to know what works and what doesn't.

2sidedmtl are very useful. What is the eta for supporting this?

[vlado]

2sidedmtl are very useful. What is the eta for supporting this?

It works already to a large extent. Using it with refraction for the base material is odd though; if it was possible I would not have allowed it even in the production renderer

Best regards,
Vlado

[tricky]

It works already to a large extent. Using it with refraction for the base material is odd though; if it was possible I would not have allowed it even in the production renderer

Best regards,
Vlado

Hi Vlado. Can you tell me the way you would create a shear curtain material? The attached image is a 10cm piece of fabric, and the fabric itself is very transparent...something like this: http://2.bp.blogspot.com/-NnDgRIwsJA...s+picasa+2.jpg

[vlado]

I would still use a 2-sided material, but I would just put a VRayColor map in the opacity of the base VRayMtl material.

Best regards,
Vlado

[Alan Iglesias]

...Can you tell me the way you would create a shear curtain material?

Tricky, I happened to have this file hanging around so I dropped your map onto the curtains and rendered it in RT/GPU.

The material is very simple - just a 2-sided material with your map as the base material's diffuse map - but it shows I think that it works pretty much as expected.

This file also represents one of the noisier scenes I have seen (darker room/exposure - light bouncing, etc), and thus the render time of just over an hour with 2x GTX580s. For your reference, the original shot was rendered at 1600x900 pixels. A couple of Titans will probably render it in ~2/3rds of the time.

I hope it is at least somewhat helpful to you...

-Alan

[tricky]

Useful test Alan. Over an hour though???? With something like that I'd hope to have some sort of render done in around 10-15 minutes using distributed render.

I guess my thoughts with RT are that it should make the scene setup a lot quicker, and then pass over to Vray for final render. Presumably, with this scene, it was reasonably interactive to set things up, so you could tell in just a few seconds if things were looking as you expected? I think until I start working with it properly, I am not going to know how it will help or hinder us. The problem I face is knowing if spending £2k on 2 x titans will be a worthwhile investment. Only real way to know is take a leap of faith.

[Alan Iglesias]

Useful test Alan. Over an hour though???? With something like that I'd hope to have some sort of render done in around 10-15 minutes using distributed render.

Yes, the scene really was a worst-case scenario for RT/GPU noise-wise. I should have worked up another file but time was short and it was ready to go. I just wanted you to see the sheer curtains with the 2-sided material in GPU.

Until faster GPUs arrive, getting rid of noise in a final high-res GI render (especially interiors) will still show the strength of the biased renderer for some time to come. On the other hand, I can note that I have also done final 3k interiors that rendered in 7 or 8 minutes compared to a 30-minute single machine biased render, but the lighting was less "challenging" and the scene less noisy. Bottom line is that it all really depends on your scene.

I guess my thoughts with RT are that it should make the scene setup a lot quicker, and then pass over to Vray for final render.
Presumably, with this scene, it was reasonably interactive to set things up, so you could tell in just a few seconds if things were looking as you expected?

Absolutely. Here's a lower-res (1k) test render that took all of 12.3 seconds to render. Noisy yes, but plenty of fast visual information regarding materials, lightiing, reflections, exposure, camera composition, etc. No more setting bias controls, no more waiting for test renders...setting up a scene has never easier, more efficient, or more fun.



...And after a quick 3-minute trip to the "watercooler", it looks like this:



The important thing here is that the overwhelming majority of the rendering time in this case was getting rid of the noise in the darker shadow areas.

*And keep in mind that I am using 2 580s, and the dual Titan set up will be ~33% faster.

I think until I start working with it properly, I am not going to know how it will help or hinder us. The problem I face is knowing if spending £2k on 2 x titans will be a worthwhile investment. Only real way to know is take a leap of faith.

Yes, there is a certain leap of faith involved for sure. My offer to work with one of your files may make make this less of a risk for you, and that is why I offered it.

The main reason I am so excited about GPU rendering is becaue it is so much dang fun, which is what I loved about 3D so much back in 1987 when I first got into it. I swear it is the coolest thing to happen to us since the 80486!

Best,

-Alan

[Morbid Angel]

back in 1987 when I first got into it.

In 1987 I was 6!

[tricky]

*And keep in mind that I am using 2 580s, and the dual Titan set up will be ~33% faster.

So is that about right then - dual Titan is only around 33% faster than a dual 580? Don't get me wrong, that is good, but £1600 for that kind of improvement doesn't really sound as good as I'd have hoped. I guess one of the big advantages of the titans is their 6GB memory.

[Alan Iglesias]

In 1987 I was 6!

Ah...thanks for making me feel so dang young, Dmitry!



-Alan

[Alan Iglesias]

So is that about right then - dual Titan is only around 33% faster than a dual 580? Don't get me wrong, that is good, but £1600 for that kind of improvement doesn't really sound as good as I'd have hoped. I guess one of the big advantages of the titans is their 6GB memory.

Yes, with the change in the core technology (Fermi to Keplar) at nVidia we unfortunately took quite a hit in ray-tracing speed, core for core. As you noted, the main excitiment about the Titan is the 6GB RAM on board, which will make GPU rendering much more viable for the folks rendering larger files.

I had hoped for more rendering speed improvement as well, but at least the Titan renders at about 1.5x faster than the 580, who's faster rendering cores seem to be obsolete - at least for the moment. This is at least going in the right direction and by the next "round" we should hopefully be at least 2x the speed of the 580 which would be the good "next step" in GPU rendering I was waiting for.

-Alan

[tricky]

Yes, with the change in the core technology (Fermi to Keplar) at nVidia we unfortunately took quite a hit in ray-tracing speed, core for core. As you noted, the main excitiment about the Titan is the 6GB RAM on board, which will make GPU rendering much more viable for the folks rendering larger files.

I had hoped for more rendering speed improvement as well, but at least the Titan renders at about 1.5x faster than the 580, whose faster rendering cores seem to be obolete - at least for the moment. This is at least going in the right direction and by the next "round" we should hopefully be at least 2x the speed of the 580 which would be the good "next step" in GPU rendering I was waiting for.

-Alan

I'm wondering if there is a chance that RT could be 'adjusted' (from a coding point of view) to make better use of Keplar?

[Alan Iglesias]

I'm wondering if there is a chance that RT could be 'adjusted' (from a coding point of view) to make better use of Keplar?

I believe Vlado mentioned that he would look into it, and certainly the latest version is rendering ~14% faster in CUDA mode than previously. If there is more that can be done, then that would be great of course.

I think it has more to do with perhaps nVidia realizing it did not need all that computing power for gaming, which of course is their main concern with these GTX boards. It might also be noted that, with the advent of the Keplar technology, power and cooling requirements relaxed significantly, which was a good thing for their marketing department, I'm sure.

Best,

-Alan