Question about PHB in output of nvidia-smi topo -m

AI & Data Science Deep Learning (Training & Inference)
1

I want to see topoloy of GPU and NIC in my GPU server, so I run command “nvidia-smi topo -m”. And in the output of “nvidia-smi topo -m”, I find that the connection of two GPU is PHB, and I am very confused with PHB. If connection of two GPUs is PHB, does it means that if a GPU want to send message to another GPU, the GPU must send message to CPU and then CPU send message to another GPU?