Training neural networks is requires extremely high speed interconnect between the humongous GPUs