Multi GPU problem

Hi, all.
I’m from Beijing. I just a few weeks ago started Eddy.
And also I’ve tried install to My new machine with 4 GPUs there.
then also Set it Eddy ENV and checked also Nvidia ENV already followed Eddy document as Instruction.
The problem is all 4 GPUs are seen at the terminal when Nuke startup, but immediately can see a below message and shut-down the Nuke in one second.
I already spent my whole day. Anyone can help me?

  • Cent 7, Nuke 11.3v4, 4 GTX 1080 Ti, Eddy 2.4, CUDA 10.1


Eddy requires P2P access between cards, and unfortunately not all NVidia GPUs support this anymore.
From our experience only the Quadro, Titan RTX and Tesla range are supporting this. Some posts suggests that other cards work as well, but it appears that driver support has been disabled for these since ( eg Geforce range ).

While P2P access was once working with a RTX 2080 TI it now seems to be disabled even with an appropriate NVLink adapter. We tested this in depth as well and couldn’t make it work, so it appears to be something NVidia is no longer supporting. NVidia provides a simpleP2P test application in their cuda framework and even this seems to be failing on previously supported cards these days ( despite working in older forum posts ).

We successfully tested Eddy on these cards configurations:
GP100, GV100, Quadro RTX6000, Quadro RTX 4000.


Thank you very much for your quick reply.
In your answer, Nvidia no longer supports the P2P of my graphics driver, which means I can never use eddy as a multi GPU environment. I’m very sorry.
It means that you already know these circumstances and limited issues, and I think you need to officially mention them somewhere. In my memory, I don’t remember seeing anything other than a single line of promotion to support multi GPU.
Well-known renderers such as the GPU-based “Red shift” appear to be working with commands that force all GPUs to be activated. We’ve activated all four 1080 Ti and we’re using it well.
I wonder if you can’t develop this kind of forced recognition option.



Hi Andy,

I feel your frustration. We were in a similar situation ourselves where a dual setup is no longer working. We chose the P2P approach as the provided infrastructure seemed appropriate at the time. We currently have no plans to add another option to support multi gpu rendering apart from the P2P approach.
Please note we also simulate across multiple cards, which invalidates other approaches that work well for rendering, but not for simulation where the intercommunication bandwidth between the cards is a lot more important.

While it doesn’t help your case, we updated the documentation in that area as well and will be in the upcoming release.