What is the NUMA situation for 4090 p2p ?

[scientific-coder]

The next tinybox (tinybox pro) will have 8 GPUs and 2 CPU, so I'm wondering about the p2p situation, presuming that each CPU will have a PCIe root.
On one hand, 8 is much better than 6, but on the other hand, I'm worried that for training tasks that could be accomplished by 6 4090 but wouldn't fit in 4, the tinybox pro could be less performant than the tinybox.
Surprisingly, it seems that p2p can work across NUMA but I wonder if this is supported by this driver and about the performance implications for training.
Could a clarification about the NUMA situation be added to the documentation ?
Thx !