RTX Pro 6000 P2P Bandwidth Slumps

This looks less like a bad benchmark and more like a PCIe topology problem wearing a GPU costume.

• –`PHB` usually means traffic is traversing the CPU/root complex, so direct GPU-to-GPU transfers may be constrained even on Gen 5 x8 links
• –NVIDIA docs note that PCIe peer-to-peer performance can be hurt by ACS, IOMMU, and routing through the root complex, all of which can crush bisection bandwidth
• –Having to disable Resizable BAR just to keep the pair stable suggests the board is already in a compromise configuration
• –The surprising part is not that P2P is slower, but that host-mediated copies can beat it when the platform cannot provide a clean peer path
• –For dual-GPU AI rigs, this is a reminder that motherboard BIOS, chipset behavior, and slot topology can matter as much as raw GPU specs