I'm trying to make blender work with 20 GPU's, however I'm able to concurenntly force only 13 of them to work together while rendering one frame(13 tiles, one gpu per tile). Is there any limitation of GPU's that can run together?

I know octane have/had 12 GPu's limitation, however i can not find any piece of information about limits in cycles.

  • 1
    $\begingroup$ I'm not aware of anything. You might get a quicker answer if you ask the devs over at devtalk.blender.org. $\endgroup$ – Sazerac Mar 18 '19 at 23:38
  • $\begingroup$ For anyone looking for informations about that, blender developers said that they are not aware of any limitations however cycles wasn't tested on that many gpu's. For blender 2.8 commit 191b8951f7af everything works fine for both CUDA and OpenCL for up to 10 gpu's. Thread link: devtalk.blender.org/t/… $\endgroup$ – PStarczewski Mar 20 '19 at 12:23
  • 2
    $\begingroup$ do you want to post a summary of that as an answer so we can point others back to this question? $\endgroup$ – Sazerac Mar 21 '19 at 23:30
  • $\begingroup$ @Sazerac on it. $\endgroup$ – PStarczewski Mar 22 '19 at 16:02

TL;DR There is no official limit of GPU's on cycles but setups bigger than 10 GPU's were never tested so far.

After not so extensive research I've contacted one of the blender developers. According to him in this thread: Blender dev forum there is no official limit on cycles. However blender 2.80 is still under development and such a huge setup was never tested before by no one. At this moment I am unable to check it myself on 20 GPU's but I've done my best with what I had.

Natively(without my software, just pure blender) I was able to render on 10 NVidia GPU's both on CUDA and OpenCL concurrently.(commit 191b8951f7af, NVidia driver 398.11, Win10). Further commits were fine for CUDA but I was unable to run OpenCL so I had to checkout from the most recent to this particular one(no idea why, but it's beta so everything is possible). That is the maximum amount of GPU's I had on one node.

Non-natively(using my software to connect to GPU's via tcp/ip) I was able to render on 13 NVidia GPU's on OpenCL(no way to render on CUDA that way) concurrently.(commit 191b8951f7af, NVidia driver 398.11, Win10). All 20 GPU's were visible though. At this point(and in the closest future as well as I have to focus on the other project for now) I cannot conduct any other tests to check whether it's blender or my software.

When I will be back on this project I will provide more information both here and on blender thread, but for now that is all I know.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.