We have a cluster with InfiniBand interconnects. I know that Distributed TF uses gRPC to handle network calls, but I was wondering if there was anything special required to gain the speed of InfiniBand connections when using the gRPC.
I think the answer is no. And technically is not Tensorflow that would need the infiniband support but more the gRPC layer.
Doing a quick query for infiniband in the github gRPC repository and receiving zero results strengthens my assumption: https://github.com/grpc?query=infiniband