27 Sep 2017
Neural networks training is a time consuming activity, the amount of computation needed is usually high even for today standards. There are two ways to reduce the time needed, use more powerful machines or use more machines.
The first approach can be achieved using dedicated hardware like GPUs or maybe FPGAs or TPUs in the future. But it can also be done by splitting the task between more general purpose hardware like the one used in cloud systems.
This document summarizes the conclusions reached after researching the use of distributed neural networks.