Training can happen asynchronously and in batch - inference needs to be low latency and have massive parallelism because it's powering millions of android devic

Training can happen asynchronously and in batch - inference needs to be low latency and have massive parallelism because it's powering millions of android devices voice to text recognition - for instance.