Network GPU Rendering

[modernleper]

We are having issues utilizing the nVidia Tesla 2070 GPUs in our network rendering nodes. Earlier in the summer we had done some preliminary testing and in the Indigo Network Slave log it would list the GPUs by name when beginning a render and we had a huge performance boost running Indigo on these machines rather than the servers without GPUs.

We have rebuilt our testing servers and are working on implementing our render farm into production, but we are now no longer getting that performance boost from the GPUs. I know we are currently running a newer version of Indigo and I'm not sure if our issue stems from that or not. Before we start a network render we go to enable GPU acceleration and we are given 2 options Cypress [Open CL] and Intel(R) Xeon (R) CPU... [OpenCL]. I don't remember these options being there in the past, but it seems that by enabling GPU acceleration we have to select one of these options and the server GPUs are forced to use OpenCL instead of CUDA. The logs on the Network Slave clearly show that it is trying to use Cypress even though it is an nVidia GPU in the render node. When we run with GPU acceleration now we take a major performance hit vs. CPU only rendering.

Am I doing something wrong? Is there a limitation sending a job from a machine with an AMD card to a server using an nVidia card? Any help

[CTZn]

Being fast, my answer will certainly be too short and naive also because the points you are rising are out of my own field of competences.

But one thing I know of that can change the list of available devices and GPU performances at once is to get and install newer drivers. Those shipping with the hardware are often outdated, and Indigo is specially demanding in this regard. Usually one can not tell by other applications, they would work flawlessly while Indigo would require the update.

Let us know if the issue was addressed that way, there is a good chance as I see it. I'll let the real guys answer to you in more details otherwise.

[modernleper]

Thanks for the quick response. However, both the server and the workstation have been built from the ground up using the most current drivers from AMD (workstation) and nVidia (server) within the past 2 weeks. We also have the latest release of Indigo Renderer, Indigo for Revit and Revit 2013 (though I don't really think those last 2 factor into the equation here).

I'm assuming it's either a setting issue or I'm just trying to do something that can't be done. I could have sworn it was working correctly using a previous version of Indigo, but I may have been mistaken.

[galinette]

I can do some test tomorrow, I have a Dual Xeon with a Tesla 2050 at work.

Question to Glare : Is it possible to use GPU nodes from a non-gpu master? (If set to non-working master, of course)

Etienne

[modernleper]

Thank you for your help and testing this issue. I have a bit more information to add. We ran a render test on 5 network slaves. The hardware specs are the exact same on all 5 except 2 have nVidia Telsa cards. When we run a render without GPU rendering enabled all 5 servers give us about 12000k samples/second based on what the renderer displays. When we turn GPU rendering on for the same job our 2 GPU servers drop to 800-900k while the 3 CPU only servers give us 1700-1800k samples/second.

Just thought the extra info might be helpful.

[JuanPabloCuervo]

I can´t find any info on the website about Network GPU Rendering support...

but i ask... does the Indigo Network Slave supports GPU ?
if so... is Not Working...

Win7x64
GTX680 306.97
HD7970 12.10

can Indigo Renderer be run in Slave Machines?

Code: Select all

Executing '-n s'
Indigo Renderer v3.4.16, Windows 64-bit build.
Licence not verified, running in free mode.
Waiting for a master.