I tried it, with NVlink. The speedup during inference is negligible. You'll probably benefit more during training.

I use 2x 3090, no nvlink though. I’d read it doesn’t help that much but not well read on any improvement.

I have not personally gotten to test things that way