FurryBall 4 - new speed compare (unbiased rendering) V-RAY, Iray, Mental ray, Arnold.


#1

Hi All,

FurryBall 4 beta is available on our web for registered users.
http://aaa-studio.cz/furryballforum/viewtopic.php?f=25&t=4139

3DS Max version will be available in Beta tommorow! (If all going well :argh: )

For our new web, we made big speed compare with other ray tracers.

http://aaa-studio.cz/FurryBall/FurryBall4compare.jpg


#2

Nice test. Furryball is coming along fine. You have used the latest and greatest Nvidia Gpu in your test setup, would be great to see the test with the latest and fastest intel cpu as well. Not that it would beat Fballs rendertimes, just to see the difference between other renderers.


#3

You are right, but it’s hard to compare.
According this page the BEST Intel Xeon cost 1,900 USD, and it’s 3x faster than i7 we used.
But for this price 1,900 USD you will have 2x Titan :cool:

For the price 1,900 USD you can have 2x Titan and FurryBall render around 2 min and CPU renderers around 15 min (3x faster instead 45 min)


#4

To be honest I think you shouldn’t do these time comparison with other engines. There are just so many tweakings and technique in each engine that it’s just unfair, and call for argument which might not end well (like your last Vray thread). The very obvious noise in your render for example.

Just show how fast you can do it and quality you can archive with it. That’s enough to gain respect. Doing it this way will only backfire on you.

Just my advise but you should look at Redshift’s approach
http://www.redshift3d.com/
They never mention any other engine themselves. The beta users are posting how happy they are with the speed and do the comparison yes, but the dev themselves stay really passive and just let the word of mouth do its thing.


#5

I don’t agree - we will put this scene to new web and everybody can test it yourself.

You are right, that there can be some tweaking, but just in few minutes. There is nice visible, that all CPU renderers has very SIMILAR times. You can’t cheat, that GPU is MUCH faster than CPU.

BTW, as you see Octane is 2x faster than FurryBall, but without some features. We tried to be
impartial.


#6

Hi, i am currently at work, so i am running a little short on time. But i promise that as soon as i get home, i will be sure to ruin your ridiculous test results. :wavey:


#7

Ehm, you cannot have unbiased or “full” GI with just 3 bounces and without caustics. What you show are results with heavy image bias. Also photon mapping is not unbiased (Mental ray).

I’d also like the 3dsmax scene, I suspect multiple CPU renderers can beat the furryball time, given the resolution.


#8

Oh, and you forgot to mention that DoF in Furryball is obviously fake. On the first render, you can see that noise in areas blurred by DoF has became splotches. It is a clear sign that DoF is applied after ray tracing calculation, and therefore it is not true ray traced DoF.

So while making my tests, for sake of correct comparison, i will be using fake DoF too.


#9

that could be just very bad JPG compression, but yeah, I saw that too


#10

Your math is OK, but we are still comparing apples & oranges here.
I look at your test and appreciate Furryball for what it does but I wouldn’t try to use it for feature film quality work for example, even if it would render that pic in 10 seconds.


#11

sorry for carpetbombing the thread, but I entirely forgot the main point I wanted to make :wink:

Please see this paper: http://pcl.intel-research.net/publications/isca319-lee.pdf (yes, it is from intel, so it may be biased, but not more than your comparisons - you are both just trying to push sales ;)). GPU is not inherently faster than CPU, it is just better for some types of tasks. Pure ray tracing seems to be in the middle between suitable and unsuitable (with complex shading being entirely in the unsuitable part of the spectrum). I wish people would just accept the fact that they will never get their magical “100×” (because I hate being asked for it every week or so ;))


#12

:thumbsup: Nice article… 3 years old, comparation i7 with 280GTX… which has 240 cuda cores, instead of Titan for example with 2688 cores… :buttrock:

We are not able to run our bench scene on this OLD card, but for example faster 650GTX with 384 cuda cores is 7,5x slower than Titan…
So 280 GTX from your article can be 10x slower than Titan…
The best Xeon is just 2,5x faster than i7 from your article and 2x more expensive than Titan…
GPU is future.
http://www.aaa-studio.cz/furrybench/benchResults4.php


#13

You misunderstood the message of the article. It does not claim CPU is faster or equally fast, it debunks the myth that GPU is/will be 100 times faster and is the only future of computing, that was often claimed at the time (and still is).

As for the 280 vs titan comparison, the best I could find is this: http://forums.cgsociety.org/showpost.php?p=7542709&postcount=9. Multiply by some factor 280 vs. 470, and you get some quite-not-10x number. But this is just empty flamewar, just please post the scene so the community can get some real results.


#14

We just released the first 3DS Max beta version :bounce: :bounce: :bounce:
Please enjoy:
http://furryball.aaa-studio.cz/furryball4Beta/index.html


#15

Again, you shouldn’t be trying to compare render times with other engines in this manner. I could demolish your mental ray time, for example, in the same scene and with higher quality - with only my Phenom 1100T. What this shows is that you don’t know the other rendering engines well enough to optimize them, and thus all your other times are irrelevant if not outright deceptions.

I think you’ve got a good product, even though (just like the rest of these renderers) your math is inherently flawed. I think if you were to apply modern physics you could further reduce your rendertimes, and then you would have a way to really compete.

To test my theory, please define a photon for us and how your scene interprets them.


#16

Coming from someone that owns vray, MR and Furryball, I too think these comparisons are crap.

Those vray and MR images are not being rendered at optimal settings, GIVE me a break! Do you take us for a bunch of noobs? It’s just insulting.

Why don’t you compare to fx shaders in maya? It’s more up your alley.


#17

Maybe I forgot mentioned, that all scene is lid ONLY by indirect (one Direct light from windows and sky also only from window).
FX Shaders will not do this… :bowdown:

BTW, here is MAX scene (with Mental Ray setup we used)
and Maya scene (With FurryBall setup)

http://furryball.aaa-studio.cz/files/Compare_indirect.zip


#18

great, thank you very much for posting the scene - now we can finally have a real discussion instead of empty speculations!

I converted the scene to a production-ready CPU renderer as best as I could. I didn’t had the displacement map, so I’ve used some procedural instead. I also had to replace the directional light with a realistic sunlight model. Here is the result in equal time to furryball:

3 bounces:
http://i.imgur.com/fyvbit6.jpg

25 bounces:
http://i.imgur.com/z3tDIsM.jpg

Notes:

  • the result is much brighter for 25 bounces (I actually had to decrease exposure by 2 stops) and is much more physically correct. It shows the huge bias of your claimed unbiased renders comparison. The scene looks weird because your walls have RGB 255 255 255 reflectance, something you never want to do.
  • The DOF is really raytraced, not done in postproduction as in furryball and probably mental ray in your renders.
  • The images was rendered with a single i7-2600 CPU, with integrated GPU. If you consider that you need some decent CPU with the titan card for GPU rendering, the price difference goes very close to the GPU cost ($1000).
  • The settings I used are almost the default ones. I am also pretty sure that with correct VRay setup you would be able to beat even these images in terms of time and quality.

CPU is not dead ;).


#19
  1. You have less bounces (FurryBall has 1 primary and 3 bounces) - nice visible on glass ball and also on scene darkness (on your second image is no much visible… and it’s little weird, that time with 3 and 25 bounces is almost same…)
  2. You have much more grain and alias compare to FurryBall. Yes, we had also similar times on CPU on this “preview” (6 min with Mental Ray) but if you need production quality, time is 44 min. BTW - On similar preview quality are FurryBall times in seconds.
  3. on FurryBall or MR is NO postproduction

#20
  1. I’ve rendered the images with a time limit, thats why the times are (more or less) the same. There is an ambiguity in what bounces are, I use a slightly different definition. Here is an image with 4 of “my bounces”. You can see, that the result is almost the same, if not better, than with 3:
  1. No. I actually have less noise on some portions in the foreground (red ball, floor). Furryball has significantly less noise in the background, but that is because the image is blurred with a postproduction DOF filter, which leads to
  1. yes, there is a postproduction DOF filter and/or interpolated GI solution in the furryball image. That is the only explanation for the blurring of the image (I assume the image uses path tracing). If the DOF was unbiased, not faked, there would be uncorrelated per-pixel noise, not weird splotches. The application of the DOF might happen in the renderer, I am not saying you used some postproduction software, but that does not change the general principle.