NVIDIA Kepler

[Karonte]

Hi All.
I love working in the gpu mode ( it's expansive!! consumes a lot of energy having gpu boxes).
And I really love Vray.
But I must say that working with gpu, it has some problems that competitors have not.
The most obvious problem is the jugged edges. Often they appear and do not go away! I tried 3 gpu competitors engines, and I must say that no one has these problems.
I think because of CUDA...i cannot find any other reason. Also the "look" of the noise is similar for all while in vray is more...how can i say?... "hard". it seems that the others have some filter....

The fact is that i have to use 2 kind of software.. one for architectural/animation (vray) and another gpu unbiased for catalogues . it's hard to have 2 material library and convert object....

I hope that vray gpu can achieve the quality /speed of competitors because I think chaos have done great job until now.

PS: in the attachment the example of the jagged edges and (small picture) a quick competitor render.

[vlado]

The most obvious problem is the jugged edges. Often they appear and do not go away!

This is because we don't have a pixel filter; this is fixed for the next update.

I think because of CUDA...i cannot find any other reason.

Nothing of the sort. It's because we currently don't have an AA filter on the image.

PS: in the attachment the example of the jagged edges and (small picture) a quick competitor render.

Can you post the full resolutions of both renders with the same materials?

Best regards,
Vlado

[cecofuli]

Vlado, pixel filter will solves the jugged edges problem for both production and RT?
Because if I want a full floating point image in VRay-Producer, with clamp output off, I have always jugged edges...
And I don't talk if I turn off sub-pixel. x3 x4 rendering time with a lot problem in th every bright area...

[vlado]

Vlado, pixel filter will solves the jugged edges problem for both production and RT?

It solves it to same extent that it solves it for other renderers...

Best regards,
Vlado

[cecofuli]

That's really great! )))

[Karonte]

Hi Vlado i made a very quick test as u told me. I have lot of deadline so...sorry for the poor scene. i'll try to make some better.

-Same scene. no Environment light.
-Same default material.
-Same light type. ( in Vray i made a box and use a VrayLightmesh, in the other software a lightmaterial).
-3.40 minutes in a single 580gtx.
-different tonemapping ( but i think doesn't matter)

Can i ask you if by pixel filtering you mean a "standard image aa" production-like? (area, etc)?
or, as cecofuli said, is it a new technique for both the RT and the production renderer?

thanks!

[Dariusz Makowski (Dadal)]

Why did u use VrayLightMesh and not VrayLight?

[Karonte]

because i think is better to make a good comparison. real light.

[Dariusz Makowski (Dadal)]

Yea the vraylight give real light, u should use the tools that are best for the job not that are matching in different programs... The RT dont have AA filter so jugged edges are normal untill next sp roll out...

[vlado]

-different tonemapping ( but i think doesn't matter)

It does, as it changes your perception of noise levels.

Can i ask you if by pixel filtering you mean a "standard image aa" production-like? (area, etc)?

Yes. Below you can see to images out of V-Ray RT GPU; one is rendered without pixel filtering, and the other is with a pixel filter. As you see, the filtered image looks much better aa-wise around the area light (they both took the same time to render). Athough the filter in this case is not one of the standard ones, we developed it specifically to deal with AA on high contrast edges.

With filtering:


Without filtering:


Best regards,
Vlado

[Karonte]

thanks Vlado for the answer!
You are always fast and kind! good work!

hem.... when will be released the update?

[3LP]

Yup, throw two of these puppies in and you have the equivalent of six GTX-580s! ...And the GPU clock freq starts at 950MHz!

The future looks better all the time...

-Alan

Hi Alan!

Could you tell us from where you have those stats ?
It looks like the new generation is 3 times faster then... is it ?

Thanks!

Stan

[super gnu]

http://www.theinquirer.net/inquirer/...70-gpu-compute



not the most reliable source. but .. hmmm.

[Alan Iglesias]

...Could you tell us from where you have those stats ?
It looks like the new generation is 3 times faster then... is it ?

The stats came from the report referenced in the first message of this thread.

Will it be three times faster? Hard to say exactly until we can do some testing. So far in my tests the number of cores has scaled on a pretty much 1:1 basis as far as rendering speed in RT/GPU. However, with the Kepler they apparently will be giving up the separate Shader Clock Speed in order to better facilitate the larger amount of on-chip cores. This is a compromise because it reduces the efficiency of each core, but perhaps it will be balanced by the overall higher GPU clock speeds.

I'll be looking forward to testing all things Kepler!

-Alan

[cecofuli]

http://www.evga.com/articles/00669/#GTX680FTW

680 and 4 GB... But I think, will cost more than 600 euro.