Page 2 of 6 FirstFirst 1234 ... LastLast
Results 11 to 20 of 60

Thread: GPU accelerated GI in 3ds max 2017

  1. #11
    Join Date
    Apr 2009
    Posts
    511

    Default

    the 980ti is more then 2,5 times faster then your current one, just to mention.

    scene complexity is ram dependent, speed is corecount dependent. and all together are clock dependent, thats the basic rule. redshift can handle more because its using the system ram as cache for example.

    so high core clock with high core count and more ram is preferable.

    and if you want to render with a gpu like with a cpu, a quadro is better, its designd for that.

  2. #12
    Join Date
    Dec 2004
    Location
    Moscow, Russia
    Posts
    682

    Default

    Kzin, Rajko in beta section posted performance table for different GPU memory size. Can you post it here? If it still actual and allowed.
    Last edited by Puppet; May 11th, 2016 at 18:00.
    Pavel Ledin

  3. #13
    Join Date
    Apr 2009
    Posts
    511

    Default

    because its beta stuff i think i am not allowed.
    and because its beta stuff, i am not sure if this would help in current mr max release?

  4. #14
    Join Date
    Mar 2015
    Posts
    11

    Default

    Quote Originally Posted by Kzin View Post
    scene complexity is ram dependent, speed is corecount dependent. and all together are clock dependent, thats the basic rule. redshift can handle more because its using the system ram as cache for example.
    so high core clock with high core count and more ram is preferable.
    So the GPU rendering of complex geometry is still a myth in the general scenario... Its the first time when I can invest some more $$$ and Im totally lost what where. The last year I rendered a MR scene with 200M on this same 660Ti2 and the RAM was filled up to 95% after some minutes... Then started to appear a message that Windows lacks memory... click OK... another time after some minutes... etc. etc. but the scene rendered successfully. So 2x xeons with 128GB is what I will get, and this 980Ti. Not sure about the cooling, will ask some genius.

    Just another question: in test scenes with more polycount (say 200K) MR freezes for forever, with displacement map slot used, and Vray renders up to 30% and closes MAX. Is it my lack of hardware, or the displacement maps are not to be used in hi-poly?
    I know the usual scenario, when artists export disp map and use low-poly with it in engines... But with high poly I cannot do?

    Instead I put displace modifier with tha same (noise or bitmap or etc.) map. This way it works fine, just needs much more geometry, to look enough good. I suppose its the same when the disp.map is used on slot - the algorhytms are calculating geometry, but with modifier it works and with slot - not.

    Imagine a spaceship, soaked in alien acid... That just a simple piece for illustration, I plan to use complex mix of procedural and semi-procedural materials. (where is the MR curvature node?!? more MR-working noises for MAX?!? I would donate for translation of BerconMaps to MR, anyone else?)

    1414awh.jpg

    If I have a rich scene with many, I mean many, objects and surfaces, I would use a procedural MR material instead of unwrapping, at least the background and not so important objects... So I would use displacement on 100 million polys. Using slot its imposible, while its OK with modifier.

    Cheerz!
    Last edited by murzilka; May 12th, 2016 at 08:47.

  5. #15
    Join Date
    Apr 2009
    Posts
    511

    Default

    Quote Originally Posted by murzilka View Post
    So the GPU rendering of complex geometry is still a myth in the general scenario... Its the first time when I can invest some more $$$ and Im totally lost what where. The last year I rendered a MR scene with 200M on this same 660Ti2 and the RAM was filled up to 95% after some minutes... Then started to appear a message that Windows lacks memory... click OK... another time after some minutes... etc. etc. but the scene rendered successfully. So 2x xeons with 128GB is what I will get, and this 980Ti. Not sure about the cooling, will ask some genius.

    Just another question: in test scenes with more polycount (say 200K) MR freezes for forever, with displacement map slot used, and Vray renders up to 30% and closes MAX. Is it my lack of hardware, or the displacement maps are not to be used in hi-poly?
    I know the usual scenario, when artists export disp map and use low-poly with it in engines... But with high poly I cannot do?

    Instead I put displace modifier with tha same (noise or bitmap or etc.) map. This way it works fine, just needs much more geometry, to look enough good. I suppose its the same when the disp.map is used on slot - the algorhytms are calculating geometry, but with modifier it works and with slot - not.

    Imagine a spaceship, soaked in alien acid... That just a simple piece for illustration, I plan to use complex mix of procedural and semi-procedural materials. (where is the MR curvature node?!? more MR-working noises for MAX?!? I would donate for translation of BerconMaps to MR, anyone else?)

    1414awh.jpg

    If I have a rich scene with many, I mean many, objects and surfaces, I would use a procedural MR material instead of unwrapping, at least the background and not so important objects... So I would use displacement on 100 million polys. Using slot its imposible, while its OK with modifier.

    Cheerz!
    for gpu and mental ray, a nvidiaARC guy should awnser this. but in general, its getting better and better for gi next.
    keep in mind, for now, just gi next is gpu accelerated. but the plan is to shift more and more parts to gpu.

    i dont know why your 200k polyscene freezes, because its not that much. but it might be the disp settings you use (people often use to extrem settings in mr). so we need more information on disp settings and input meshes and what kind of dispmaps you use, resolution for example.

    and using a dispmap on a 100mio polymesh might be to much. depends on your settings, but with detail ones you can easily create several billion polygons. you would need alot of ram todo this.

  6. #16
    Join Date
    Dec 2004
    Location
    Marina Del Rey, California
    Posts
    4,143

    Default

    I'm a bit confused myself on what is being discussed. So perhaps my clarification will lead to further clarifications.

    The GPU usage for GI that we have now in the UI in 3ds max 2017 was not available a year ago. So you must be talking about using MAX script to enable the older gi gpu string option? Please confirm. That was an older mechanism, which we re-implemented significantly into the latest GI, that we now in the UI call Indirect Diffuse (GI) which you can turn on in the 2017 3ds max as you have shown in your screen captures. (Internal name is GI Next if you see that term thrown around.) So you shouldn't be referencing the older mechanism when talking about GPU memory. The new mechanism doesn't require use of textures for example in GPU memory. It also handles motion blur, depth-of-field, and has several other production related advantages.

    We can talk about the GPU acceleration in the released 3ds max 2017 here. The version of mental ray in 3ds max 2017 is indeed 3.14,. However, it is not exactly the same patch level as is in the Maya private beta. One of the issues we addressed in a patch was accommodating smaller memory GPU cards, and we are checking to see how the patch levels relate to what is in released 3ds max.

    Let's do some step by step analysis. First try a scene with roughly 500K polys, and report on whether you see GPU acceleration and how much. Then try 1M, 5M, 10M, etc. See where it runs into memory limit.

    Exactly how many physical cores are in your CPU(s)?

    In my tests with some of the most recent patches, even with 28 total cores, but using M6000, I see roughly 4-5x speed up using GPU on a 1M poly scene, and high quality (2-5), 5 diffuse bounce depth GI. I have scenes that vary from 2x to 20x speedup depending on how demanding I set the indirect diffuse (GI) parameters and the scene. This is actually what I used to demo on the NAB show floor, as well as at GTC, our tech conference specializing in GPU.

    We highly recommend pro GPU cards, but if you stick with the consumer card, then we just announced the 10-series from the latest Pascal (P) generation last Friday. Of course, we will have to coordinate the GPU acceleration with the appropriate drivers that match those cards. But note, you will typically get more GPU memory on pro cards. For example, we now have a 24G M6000.
    Last edited by bart; May 12th, 2016 at 18:54.
    Barton Gawboy

  7. #17
    Join Date
    Mar 2015
    Posts
    11

    Default

    Thank you very much, now I see better

    Quote Originally Posted by bart View Post
    So you must be talking about using MAX script to enable the older gi gpu string option? Please confirm.
    I used this script in the past, its in the middle of this page, and gives me about 2 to 2.5 times faster AO rendering with this 660Ti and also with my laptop GeForce GT 550M which have some CUDA too.
    https://blog.mentalray.com/2014/07/0...ine-prototype/

    Exactly how many physical cores are in your CPU(s)?
    Both are 4 cores / 8 threads (i7 4771K and i7 2670QM)

    Quote Originally Posted by Kzin
    dont know why your 200k polyscene freezes, because its not that much
    Well, it freezes at 25K too... Here you can test http://www.filehosting.org/file/deta...4/2017disp.max

    The panel below is with displacement map 3%, the right panel is with modifier. The MR settings are by default, nothing changed.
    By the way just now I tryed with Scanline renderer and it freezes too. Not maybe MR problem. Vray renders displacement over 2M object, crashes max with disp on 8k object

    P.S. 100M with displacement done by modifiers is not a problem and this is a solution for me.

    Cheerz!
    Last edited by murzilka; May 13th, 2016 at 15:33.

  8. #18
    Join Date
    Dec 2004
    Location
    Marina Del Rey, California
    Posts
    4,143

    Default

    Please note that GPU for AO is different than and separate from GPU for GI.

    Also that script is for the old GI with GPU that is no longer in mental ray in 3ds max 2017. The new version actually in the 3ds max UI now should be better in many ways.

    I'm thinking that there should be some GPU-accelerated speedup if you have 4 physical cores, and that GPU. But we may need to investigate further. Again, would like to proceed with the step-by-step approach, examining the reports in the Render Message window.
    Barton Gawboy

  9. #19
    Join Date
    Jun 2009
    Location
    Sofia, Bulgaria
    Posts
    556

    Default

    There are a few big things i'd like to bring up to ur attention.

    GI Next in max has huge workflow related oversights - first and foremost it completely disregards any skylights in the scene and instead directly samples the environment map/color.

    This behavior is inconsistent with how other mr environment lighting solutions work - Final Gather and IBL both use the Skylight as a mediator, while at the same time the skylight provides additional control and functionality.

    As it is now, in order to get separate reflection and lighting environments u have to use color overrides/switchers, while the familiar "Skylight" workflow allows having the reflection environment loaded in the Environment panel and a separate map for lighting inside the skylight. Sadly the fact is that GI next ignores the "Final Gather" (which u normally use for environment lighting) part of the switcher and again uses the "Environment" rays.

    This thing will also causes every existing scene that took advantage of the mentioned workflows to render "incorrectly" if u simply enable the Indirect Diffuse (GI) option. The skylight in the older scenes provides an easy way to have separate environment/lighting and increase/decrease the environment lighting intensity, but now it doesn't work with GI Next.

    Consider also that in the Skylights & Environment Lighting panel, one of the "Skylight Modes" is listed as "Skylight Illumination from Indirect (GI or FG) also leading the users to believe that the new GI works with skylights, while in fact it doesn't.

    It's also important to add that using portal lights with GI next (in the classical scenario where u put them on windows, etc.) is a big "No" since they actually block the environment lighting contribution.

    The saddest part is that none of this is covered by existing documentation.
    Last edited by moulder6; May 15th, 2016 at 17:40.

  10. #20
    Join Date
    Dec 2004
    Location
    Marina Del Rey, California
    Posts
    4,143

    Default

    It should work. GI Next should behave the same as FG in that respect. It should be a higher quality replacement for FG.

    In fact, if using environment lighting with FG now, I see similar issues. I set the environment to red to clarify the difference between it and the bluish skylight.

    So, something about how this translates may have changed.

    This is in addition to us currently checking out FG specific switching going on in the max textured skylight and environment shaders.
    Barton Gawboy

Posting Permissions

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