I am setting up my GPU server with 2 GPUs. When I got my server from manufacture, these GPUs were installed into the different CPUs’ sockets so there is no Peer-to-Peer access between 2 GPUs. I think P2P access has more advantages for accelerating my deep learning programs because of direct memory access between 2 GPUs. So, I am just wondering why my manufacture install my GPUs without P2P capability. Can anyone explain if there is any advantage of no P2P access between 2 GPUs?