Page 2 of 4 FirstFirst 1234 LastLast
Results 11 to 20 of 32

Thread: 3ds Max 2015 and mental ray GI GPU prototype

  1. #11
    Join Date
    Apr 2009
    Posts
    511

    Default

    to slow?
    we talking about a brute force gi solution and noisefree renders for interiors.
    one of my tests was the classroom scene, rendered in 1080p with phys sun/sky for lighting only, with above settings, completely shaded with mia and really high local glossy samples (ranges from 128 to 2048 ), mia because of some mila sampling problems, and rendertime was 5 hours on a i7 hexcore and a gtx580 with 3gb.
    no portal lights, no other lights to illuminate corners to solve low gi bounce values because brute force is getting slow like hell with every additionell bounce. i used 8 bounces and the rendering was nearly noisefree (some noise from mia, but thats hard to solve and will be kicked when we have bsdf mila) with phys sky and sun only. from some more tests in the scene and how i had to use mia for glossy stuff and unified values because of that, i am pretty sure the rendertime with bsdf mila can be under 4 hours, not to mention the improvements in newgi which will lower it even more (had to use presample density of 2 for example).

    some comparision from other renderers, for example, octane, runs on gpu alone, took 8 hours, arnold took 9 hours on cpu, but with alot more noise and much darker because i cant use 8 diffuse bounces, arnold will freak out because of the additionell noise (and i used lights in the windows to get in more direct light to solve it better) and mental ray with irradience particles in brute force mode took 12 hours, also here with much more noise.

    the filter of new gi is not like the filter of fg, it does something different. the new gi is always a brute force method. and you current rendertime with 1 hour is a joke compared to other current brute force solutions.
    the density is for translation from cpu stuff to gpu shaders stuff, you can try if 0.2 is enough, might be possible for your scene because its simplier.

    another question, are you using it together with fg? because newgi cant solve spec gi contribution for the moment fg takes over when necessary. but because you are using other lights in your scene it might be possible you can deactivate fg and use newgi alone, that should save some computation time.

  2. #12
    Join Date
    Dec 2005
    Location
    Wherever The Computer Says
    Posts
    2,853

    Default

    Quote Originally Posted by thorsten hartmann View Post
    Hi Remydrh,

    GI GPU is the same way as irradiance particle. You need only a mr_sky or a OpenExr in the Environment. If i have IBL on, was the image to bright. The Render resolution is 1920x2444. All on my Website. I must set the Filtersize to 20, because lower values give me cloud on the ceiling.
    It's not really the same as IP. However, there was a bug at some point with IBL + GI that caused some things to be overbright. It has since been fixed. The version in Max might be lagging behind some I think.
    "Don't argue with an idiot, they will drag you down to their level and beat you over the head with experience."

  3. #13
    Join Date
    Apr 2009
    Posts
    759

    Default

    Ah, good to know. Thanks for the infos. Can you confirm, if i set presample depth to 4, need the GPU to much rendertime and the CPU must wait. Sometime give it a crash with a other commercial scene.

  4. #14
    Join Date
    Apr 2009
    Posts
    759

    Default

    Hi Kzin,


    for me is important a good final quality and speed. For me is the quality enough, if i use GI GPU = ON ( not diffuse mode) + FG. I have in the testscene glaswalls and this need a complex lightsolution. I use here 2 x Portal light, mr_sun, 6 IES-Lights as area lights. The Unified Sampling are High-Settings, and i need only for a Big resolution max 1 - 2 hours. I think this is good. 1080p in 5 hours is not profitable.

    PS: i have read in the docuemantion, FG must be allways on!

    Current limitations
    For GI GPU to be effective, finalgather must be turned on. Some features are not yet supported: distorting lens shaders, motion blur, particles, volume shaders, camera clipping planes, progressive rendering. There is only limited support for scattering shaders, emissive materials, and hair rendering.
    Last edited by thorsten hartmann; July 13th, 2014 at 13:55.

  5. #15
    Join Date
    Apr 2009
    Posts
    511

    Default

    fg on is because of the "no specs for newgi" for the current version, its a limitation, fg takes control for indirect spec rays (and because the users should not miss the spec thing and also to go over problems because of glossy transmission shaders which can cause problems when using indirect light only). but because you have 6 direct lights which illuminating areas behind or inside glasobjects it might be possible you dont need fg on. would be worth to test because rendering should be faster without fg on.

    2 portal lights, have you tryed to render without them? shadow sampling for these two lights could be more costly than using only newgi for ibl.

  6. #16
    Join Date
    Apr 2009
    Posts
    759

    Default

    Ah a light in my brain. ok at this have GI GPU so many limits, that i understand something wrong. Thank you for the info. But 5 hours for a 1920x1080 canĀ“t no client pay me.

    yes i try without portal light, but i need more light inside the room. What the most 3D-Artist forget is, that a real Photograph use Mirrors and extra Lights for a good interior Images. Very important, if your project is a design study. I use Portal Lights to have more light energy in the room.

  7. #17
    Join Date
    Apr 2009
    Posts
    511

    Default

    Quote Originally Posted by thorsten hartmann View Post

    that a real Photograph use Mirrors and extra Lights for a good interior Images. Very important, if your project is a design study.
    yep, using arealights for this, but light emitting surfaces would be better for newgi (light emitting planes with hdr's would be the best solution). not supported yet, but pretty sure it will be a future feature.

  8. #18

    Default

    Okay, here goes my feedback.

    It is incredibly slow compared to competition, very limited, and GPU based.

    Seriously, this new GPU GI performs a lot worse on high end GPU (GTX680) than Vray's Irradiance Cache does on mid-end CPU (i7 2600k). Add also memory and feature limitations to that, and you quickly realize it is not even remotely capable of any competition. Let alone force mode, which is just terribly slow (again on high end GPU) compared to let's say Vray's CPU based brute force.

    What's even worse, is that it seems this new GI is intentionally crippled when running in CPU mode. The CPU utilization was terrible, and it took ages to even start rendering. Nowadays, when interactive rendering is knocking on the door, waiting over a minute for first bucket to show in VFB in extremely trivial scene is just unacceptable.

    When i tried Maya 2015 retail, new GI was terribly slow when in CPU mode only (cuda devices string set to 0), and CPU utilization in task manager showed poor usage. Then service pack came, claiming CPU utilization was fixed. Now guess what... new GI is EXACTLY as slow as it was before the service pack, but CPU utilization is now 100%. That shows you simply just covered up the fact it is intentionally crippled when running on CPU only.

    I can imagine why though... it would be a real shame, if GPU making company (nVidia) showed, that their new, primarily GPU-based feature runs better on CPU. Luckily enough, this cover up does not do the trick, because thanks to the presence of other CPU based renderers, people can quickly test and measure that mid-end CPU easily outperforms high-end GPU in both interpolated and brute force GI solutions, which easily puts GPUs in shame.

    So, to provide feedback, as you wanted, here are my suggestion:

    Make it primarily CPU
    Remove all the limitations
    Make it at least 5 times faster, to bring it on par with competition
    Remove excruciatingly long precalculation times, which make any kind of fast iterative process a hell.
    Last edited by rawalanche; July 17th, 2014 at 15:56.

  9. #19
    Join Date
    Apr 2009
    Posts
    511

    Default

    it would be great if you can make some example renders to proof what you are writing here. show the images from your test of newgi and vray brute force mode and explain what settings you used. i think it would help the devs to find the problems.

  10. #20

    Default

    I have wasted way too much effort posting tests on this forum just to be ignored or stopped by wall of arrogance. Tests are always subjective because everyone uses different hardware.

    There's nothing hard to understand about making it faster and less limited.

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •