Best Multi GPU method - Omniverse Farm VS PCIe Bifurcation?

I have 4 identical PCs each with their own RTX 3080s + Ryzen 12 core CPUs. I would love to occasionally get the GPUs to work together to increase the real-time viewport performance in Unreal/Omniverse.

Would it be better to use PCIe bifurcation (x16 converted to two x8) for two GPUs via NVlink within one motherboard OR the new Omniverse Farm over a LAN? It seems like the pcie x8 bandwidth would be much wider than even a 10Gig LAN network. Then again how large (how many GPUs) does Nvidia Farm allow us to scale? With bifurcation you usually are limited to 4 pcie devices (2 GPUs) but would Nvidia Farm would allow you to add many more than two GPUs?

All things considered, which method will perform better in the viewport in real-time? Thoughts anyone?

1 Like

Farm all the way, it is super flexible and so far no Nvlink support. With Farm you can scale it with a few clicks across all the computers in minutes. No fancy setup. I run 4gpus , 2 gpu, 1 gpu, 1gpu. Depending on the size of the projects and tasks farm picks them up and spreads the work across all the pc.

You want multiple for collaboration, but if you are doing rendering your limit won’t be the 8x pcie lanes.