Announcement

Collapse
No announcement yet.

How to render extreme large renderings per GPU? (speed issue)

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • How to render extreme large renderings per GPU? (speed issue)

    Hi,

    in the past I used LC+IM and was able to render large images within a good time frame by undersampling the IM. Now, LC+BF on GPU is nice for the daily work, quality and speed are in a good range. But if I need very large renderings than the render times strong grow. So, I get a standard panorama of 12000x6000 in 2h (2x2080ti). It's quite long. But I will need it in stereo and I suppose so the render time will be 4h.

    I miss the possibility to undersample. Is there a trick to save render time for large outputs?

    For the future could the GPU pipeline get an option to undersample and interpolate the calculation?
    www.simulacrum.de ... visualization for designer and architects

  • #2
    If an image is four times bigger it does take much longer. There is no shortcut.
    Just make sure you dont ran out of Vram.

    Except you use some kind of ai upressing like Topaz.
    https://linktr.ee/cg_oglu
    Ryzen 5950, Geforce 3060, 128GB ram

    Comment


    • #3
      +1 On Topaz AI Gigapixel, it is quite helpful in these situations
      could end up rendering in half of that resolution and getting very good quality compared to the native resolution

      About rendering speed, 12000x6000 is a quite massiver resolution honestly
      2 hours doesn't seem that bad here, my recommendation is using Bucket mode.. Max subdivs of 50 min subdivs 1 and noise threshold of .02
      LC default settings, just increase subdivs to 3000 or so

      You can also experiemnet with higher noise threshold and using the denoiser

      Best,
      Muhammed
      Muhammed Hamed
      V-Ray GPU product specialist


      chaos.com

      Comment


      • #4
        Thank you for your hints.
        www.simulacrum.de ... visualization for designer and architects

        Comment


        • #5
          +1 for gigapixel AI
          its been a life saver on a current job that they reckon they need 25k x 15k images....lol ok then guys ill just upres cos at 25k the geometry doesnt stand up at all..

          Comment


          • #6
            If I use the V-Ray denoiser than I get noisy edges and this edges are bad looking after upscaling per AI Gigapixel. OK, I could raise the noise threshold, but the plain surfaces are looking quite good. I wished I could focus the calculation to the edges. I ask me could it not be better to calculate a higher resolution with some noise and don't upscale it?

            Or is it better to use the NVIDIA denoiser?
            www.simulacrum.de ... visualization for designer and architects

            Comment


            • #7
              You need to raise the noise threshold. If you need high quality Denoising does work only for near clean images.
              For high quality i would let the image cook for 3/4 of the time. If i need decent quality i would go for half render time.
              https://linktr.ee/cg_oglu
              Ryzen 5950, Geforce 3060, 128GB ram

              Comment


              • #8
                I did some more tests and got:
                * 7000x3500
                * max subdivs 32
                * noise threshold 0,025
                --> 30min render time and upscaled to 14000x7000 looks quite good. But this is the limit, maybe I increase the quality to 0,02/50.

                Doe's someone try to upscale a stereo pano? I hope the upscale process doesn't create different artefacts for left and right view. It will be my next test.

                www.simulacrum.de ... visualization for designer and architects

                Comment


                • #9
                  2hrs for a 12k image is FAST. In automotive stills work 6000 x 4000 these days is small, and for a final, high quality image with DOF, you can assume itll take 4hrs+ for an interior where thr frame is filled with pixels. Id expect around an 8hr render (overnight) for a 12k render, where the renderable elements filled the frame, realistic shaders with surface imperfections and DOF. Id be very happy if I got a 12k final quality image out in 2hrs.

                  Obviously experiences will be different to someone with a stack of RTX GPUs or a 3990x cpu with distributed rendering on, so its all relative.
                  Website
                  https://mangobeard.com/
                  Behance
                  https://www.behance.net/seandunderdale

                  Comment


                  • #10
                    I ask me - is on-demand mipmapping slower than the full res mode? Should I try to use the full res mode?
                    www.simulacrum.de ... visualization for designer and architects

                    Comment


                    • #11
                      If you have the RAM to do so, Id say yes. You might get some slow down, but you'll get best texture quality, Id imagine. I've not done any like for like tests, as I render with full-res as standard since one of our heads of 3d doesnt trust the mip mapping.
                      Website
                      https://mangobeard.com/
                      Behance
                      https://www.behance.net/seandunderdale

                      Comment


                      • #12
                        You mean slow down by full res texture? I supposed so that the ondemand mid-mapping could slow down the rendering by loading/unloading parts of the texture.
                        www.simulacrum.de ... visualization for designer and architects

                        Comment


                        • #13
                          Originally posted by oglu View Post
                          If an image is four times bigger it does take much longer. There is no shortcut.
                          Yep. Sometimes it's easy to forget that a 2x larger image resolution, is actually 4x the area, so the render will take 4x longer to render 2x larger resolution.
                          www.DanielBuck.net - www.My46Willys.com - www.33Chevy.net - www.DNSFail.com

                          Comment


                          • #14
                            Originally posted by Micha View Post
                            I ask me - is on-demand mipmapping slower than the full res mode? Should I try to use the full res mode?
                            Only use mipmapped textures when you run out of memory, I use full Res texture pretty much all the time on my Titan cards
                            And if you use mipmapping, you will need to be extra careful as it can introduce issues(it might be more reliable now, but still)
                            Role of thumb usually is to make things easier on Vray, so it makes your life easier
                            Muhammed Hamed
                            V-Ray GPU product specialist


                            chaos.com

                            Comment

                            Working...
                            X