[aikinai]

Google's infrastructure gives it an extra boost, but even the single machine version can beat the distributed version 25% of the time.[1]

[1] https://twitter.com/demishassabis/status/708489093676568576

[wheresmypasswd]

The nice thing about most of this deep learning stuff is that you can use a million machine hours to train your model, and almost no time to make an evaluation. So the single machine version has all of the pattern recognition given to it by the cluster, but a few ply less tree search depth.

So to me this underscores the relative importance of the deep learning model vs the tree search.