Dual PCIe CPU Slots vs Dual PCIe (CPU and Chipset)

[Danik-droid]

I asked this here too https://www.reddit.com/r/LocalLLaMA/comments/1n4p27l/dual_pcie_cpu_slots_vs_dual_pcie_cpu_and_chipset/

So, for text generation purposes, when using two GPU cards (mainly Vulkan and a bit of ROCm), does it matter whether both cards are connected directly to the CPU?

What's the performance difference (because theoretically there should be one)?

[ExtReMLapin]

From memory (no pun here), it matters as @JohannesGaessler added support for P2P some time ago

Edit : #2470

[JohannesGaessler]

Peer access in that context only refers to NVIDIA GPUs using CUDA, I don't know whether or not it's relevant for HIP/Vulkan.

[ExtReMLapin]

Imagine if nvidia blocked p2p on cuda (without driver hack) but not on vulkan