Just wondering whether the Jetson Nano is capable of GPU direct RDMA or not?
GPUDirect RDMA is only supported on Jetson AGX Xavier, due to HW features only present on Xavier. L4T r32 supports multiple different Jetson devices using the same kernel, which is why nv-p2p.h is present even on a system that does not support the RDMA feature.
I am assuming that this means that the Jetson Nano is not IO-coherent (whereas Xavier is). So Zero Copy Memory is uncached ?
Jetson Nano does not have IO coherency, but Jetson AGX Xavier does. I am not fully familiar with the caching attributes that are used for the different types of CUDA memory allocation on different systems; someone from the CUDA team would have to provide those details, assuming they are something that we document.