ESXi: Optimal physical topology for multi-vGPU VMs?

Virtual GPU Forums General Discussion
1

Hi there,

this is a very specific question, but I think this is the only place that has a chance of getting an answer :-)

On an ESXi host, we want to use 4 RTX Pro 6000 Blackwell GPUs (mainly for inference).

The host has 2 NUMA nodes.

We want to use vGPU for better flexibility, but for now we are using only “full” vGPU instances (vGPU frame buffer size = physical GPU frame buffer size).

What would be the most efficient topology for this scenario?

  1. Connect all (4) physical GPUs to the same NUMA node
  2. Evenly split all physical GPUs (2, 2)

We understood that P2P functionality is not available (at least not via PCIe).

Is UVM performance significantly affected by NUMA placement?

Best