A GPU did the actual math in 2.6 milliseconds. Then it spent 23 milliseconds just moving the answer around.

Just think about it. The math took 2.6. The moving took 23. So almost ninety percent of the time wasn't computing anything. It was just carrying numbers back and forth.

This is from a paper about k-means. It's a clustering algorithm. If that means nothing to you, that's fine. The important part is that it shows up in a lot of modern AI systems, so it's worth knowing about.

k-means is old. 1982 old. One version goes back to 1967. You learn it in week two of a data course and then forget it.