Multiple RTX 3090 - P2P driver, NVLink or what can be done?

This driver is a bit of a game changer along with the PEX88096 class hardware dropping to $300 making it viable. All of a sudden a gaggle of 3090s or 4090s can be fully exploited. I shall want to look into selling off my 3090 NVLink bridge... its 54GB/s iirc bandwidth is almost matched by gen 4 32GB/s which can be had apparently with a pcie switch topology now. This driver and an affordable gen 4 pcie x16 switch is making affordable inference dramatically more possible. It seems like all the ingredients are there and I should be able to do insane stuff like 10x or even 20x 5060ti in my x399 threadripper. Each PEX card becomes its own GPU-NUMA domain of either 5 or 10 cards each. maybe i stay modest and run 8x5060Ti across two of those suckers to give them full P2P bandwidth. Holy hell if that works as well as it sounds like it could, you're getting a 5090 worth of compute (and twice the VRAM) out of each group of 4 of them. If x8 gen 4 is sufficient to feed them for parallelism it'd be even better, 8x5060Ti in each cluster: That's 128GB and twice a 5090's worth of compute roughly. X99 systems apparently can do dual x16 slots for something similar, but those are less exciting being on gen 3. I will want to work out whether that is possible. it would be hilariously awesome because x99 is dirt dirt cheap, and if you can unlock nearly full inference power off a single gen 4 pcie card in a rig like that it will be truly amazing. Next step is clearly raspberry pi 5, then.