Ijaw Dictionary Online

How Automobiles Work

In this video we will compare non-distributed training (Old Way) to distributed training (Engine ML) The videos are playing 10x faster than real time These two experiments are using the exact same model and dataset The model on the left is training on 1 GPU The model on the right is training on 128 GPUs The dataset is divided and across multiple machines and the model is trained with synchronous gradient descent The Engine ML experiment just completed its fifth iteration through the entire dataset The 1 GPU experiment isn’t even 10% done with its first iteration It will take almost 5 hours to get to iteration 5 Total experiment time (60 epochs)
Old Way – 2.5 Days
Engine ML – 30 Minutes What would you do with that extra time?

Leave a Reply

Your email address will not be published. Required fields are marked *