Tensorflow with MVAPICH2/OpenMPI

Hi there,

Does it make sense to compile tensorflow for jetson tx2 with MPI support?
I am running distributed inference on two jetsons and observe significant network delay that reduces performance gain to nothing.
There are a couple of optimizations available in tensorflow to speed up network communication but it seems nothing is directly applicable to Jetsons? I’ve tried ‘grpc+verbs’ and ‘grpc+gdr’ so far.

Am I missing something?
Please advice.

Hi,

We recommended to use TensorRT instead.

Jetson and TensorRT are our solution for fast inferencing.
There are too much implementation on TensorFlow targets for training and easy developing.
It may not be optimal to run inferencing on the embedded platform.

More, MPI and horovod also targets for multiple desktop GPU, not good for integrated GPU.
Thanks.