We are pleased to announce that we have added an additional communication path to TensorFlow that allows Tensors to be exchanged using MPI-based send() and recieve() routines when running distributed TensorFlow. The used MPI implementation will pick the most optimal path available between the two communication points allowing the user to take advantage of high performance networks such as Infiniband. These types of networking infrastructures are commonly found on super computers and high performance computing clusters.
How can we help?
Reach out – we’d love to hear about how we can help.