Announcement

Collapse
No announcement yet.

GPU-Oriented PCIe Expansion Cluster ?

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

  • GPU-Oriented PCIe Expansion Cluster ?

    Let's say i want to to add a lot of GPU to one motherboard and use this kind of product that lets me install 4 gpus on 1 PCIe slot

    http://amfeltec.com/products/gpu-oriented-cluster/

    Click image for larger version

Name:	SITE_1.jpg
Views:	282
Size:	2.88 MB
ID:	963424

    I was wondering if the performance would be the same as installing the 4 gpus directly on my motherboard ?

    My motherboard is an ASUS RAMPAGE V edition 10 and i have 4 PCIe 3.0. So i was waondering if there's was a limitation on a much GPU you can plug on 1 motherboard ?

    Amfeltec claims that with their cluster you can hook up, up to 16 gpus on 1 machine.

    Can anyone please shed some light on how this work

    Thank you.
    Last edited by walee; 09-08-2017, 05:32 AM.
    ____________________

    WALEE / www.walee.com

  • #2
    We have not tested those, so we can't tell. Usually, esepcially with higher rays per pixel / ray bundle size, V-Ray GPU is not very PCIe bandwitdh demanding.

    Best,
    Blago.
    V-Ray fan.
    Looking busy around GPUs ...
    RTX ON

    Comment


    • #3
      This looks really interesting. If you increased RPP and bundle size would there be a bottleneck anywhere else?

      Comment


      • #4
        Originally posted by jironomo View Post
        This looks really interesting. If you increased RPP and bundle size would there be a bottleneck anywhere else?
        I'd suggest looking at the benchmarks (https://benchmark.chaosgroup.com/gpu). You'll notice that systems with 8 cards is rendering 1 second faster than a system with 4. There is speculation in another thread of what is causing the lack of productivity above 4 cards. Other users have speculated it might be a bios/motherboard supporting that number of cards. I haven't heard anything from Chaos Group of what the issue would be, but those results steered me away from building anything with more than 4 cards per system.

        Comment


        • #5
          Originally posted by KhyeKading View Post

          I'd suggest looking at the benchmarks (https://benchmark.chaosgroup.com/gpu). You'll notice that systems with 8 cards is rendering 1 second faster than a system with 4. There is speculation in another thread of what is causing the lack of productivity above 4 cards. Other users have speculated it might be a bios/motherboard supporting that number of cards. I haven't heard anything from Chaos Group of what the issue would be, but those results steered me away from building anything with more than 4 cards per system.
          I personally believe at this stage that the Benchmark software is currently the main *suspect* !

          Let's see
          Jez

          ------------------------------------
          3DS Max 2023.3.4 | V-Ray 6.10.08 | Phoenix FD 4.40.00 | PD Player 64 1.0.7.32 | Forest Pack Pro 8.2.2 | RailClone 6.1.3
          Windows 11 Pro 22H2 | NVidia Drivers 535.98 (Game Drivers)

          Asus X299 Sage (Bios 4001), i9-7980xe, 128Gb, 1TB m.2 OS, 2 x NVidia RTX 3090 FE
          ---- Updated 06/09/23 -------

          Comment


          • #6
            Originally posted by JezUK View Post

            I personally believe at this stage that the Benchmark software is currently the main *suspect* !

            Let's see
            Thats correct

            Best,
            Blago.
            V-Ray fan.
            Looking busy around GPUs ...
            RTX ON

            Comment


            • #7
              Originally posted by savage309 View Post

              Thats correct

              Best,
              Blago.
              And that was sort of my point when discussing a non linear performance increase in the gpu render node thread

              when I tested (rank 16) with 5 gpus and got 18.876, then (rank 9) with 6 gpus and got 17.768 sec, then (rank 7) with 7 gpus and got 17.518 sec, that is not linear performance

              but if I'm rendering a 4k frame with heavy refractions, I am super happy to have those extra gpus because when working on the noise reduction step of the render they are all contributing to reducing render time during that step.

              Comment


              • #8
                Originally posted by walee View Post
                Let's say i want to to add a lot of GPU to one motherboard and use this kind of product that lets me install 4 gpus on 1 PCIe slot

                http://amfeltec.com/products/gpu-oriented-cluster/

                Click image for larger version  Name:	SITE_1.jpg Views:	1 Size:	2.88 MB ID:	963424

                I was wondering if the performance would be the same as installing the 4 gpus directly on my motherboard ?

                My motherboard is an ASUS RAMPAGE V edition 10 and i have 4 PCIe 3.0. So i was waondering if there's was a limitation on a much GPU you can plug on 1 motherboard ?

                Amfeltec claims that with their cluster you can hook up, up to 16 gpus on 1 machine.

                Can anyone please shed some light on how this work

                Thank you.
                I asked amfeltec tech at what gen and lane speed that runs at ( remember that pcie generation is a factor to consider along with lane speed...gen 3 runs faster then gen 2 at same lane speed) and I decided to go a different route.

                plus I wasn't sure if they would simultaneously transfer or if it transfers to one card then transfers to card 2 etc...too many unknowns. I remember reading in a different thread the performance of that splitter was not what the builder expected for their renderer.

                it still worked ok, but I felt there were better ways to build for my needs...plus I thought it was expensive.

                the guys using this splitter were building 20 gpu rigs due to the renderer licensing.

                Comment

                Working...
                X