It seems to really depend how you use Rayon, and the size of the data. The overhead of setting up the parallelism can't exceed the savings basically. Wherever i

It seems to really depend how you use Rayon, and the size of the data. The overhead of setting up the parallelism can't exceed the savings basically. Wherever it's possible to use ndarray broadcasting (vectorization), I do that instead. But I do think Rayon and the Rust compiler have gotten better about being smart, and I'm super impressed with the speed.