It's much faster in cases that involve significant processing. Where possible query execution is vectorized and runs in parallel on as many cores as you make available.

There are some examples in the article cited by dang: https://news.ycombinator.com/item?id=20163017