Why this trains so much faster.

Video transcript

The old way reads one word at a time. Word three must wait for word two, which waits for word one. It is a long chain.

In the Transformer, every word connects directly to every other word. No chain at all.

So all the words can be worked on at the same moment? Then a GPU can do them side by side. What does that do to the steps in a row?

One step. The old way needed a step for each word in between. Here, distance does not matter, and the whole sentence runs in parallel.

That is why the paper's big model trained in about three and a half days on eight GPUs, far less than older systems needed.