Soon to be Render Farm....

[RyanW17]

So, we’re (Work) are looking at putting together a GPU based ‘render farm’ to increase throughout in comparison to our current CPU based solution. But I’m trying to understand it a bit more and get some feedback from you guys. We’re looking at NVIDIA T4s and NVIDIA V100s for these.

 

Video processing/transcoding will be done by ffmpeg which supports HWaccel natively (As long as it’s recompiled with the drivers). I’m just wondering what kind of expected throughput would we get with the below  4x Linux VMs each with a T4/V100 in comparison to let’s say 4 VMs with 8/16 Cores each (Don’t know exactly Xeon mode but it’s a fairly new platform).

 

from what I’ve been reading so far the main reason why GPUs aren’t used is because they don’t have enough onboard memory for these tasks however we don’t plan on processing huge video files it’s more the sheer volume of assets being passed through the system,

 

any suggestions/input would be great! 

[djlithium]

Hey Ryan.
We have been dealing a bit with the older P104s (which is a mining card, I know) in tests for GPU based 3D rendering and while super slow, the T4s may be the way to go. 
With respect to the size of the assets, video isn't really impacted by this and since the T4s are 16GB each, we have been looking at them for that reason actually as the GPU rendering requirements for our customers are starting to push into that territory, but, the other thing to factor in is the wattage draw. At 70 or so watts compared to a 180w draw for a GTX-1080 (correct me if I'm wrong) you are likely best off going the T4 route. however if you want to test this I may be of assistance.
PM me for details.

Cheers.
 

K.