These are the most interesting bits for me:

> Rust is the fastest CPU language. This is not terribly surprising, as it has a mature compiler, and its default behaviour of unboxing everything is exactly what you need for this program.

> What is not visible from the above table is that most of the implementations were significantly slower in their original formulation. Only Futhark, MPL, and Rust are essentially unchanged from their first straightforward implementation. For the others, most of the performance comes down to various low-level tweaks, in particular avoiding boxing and allocations. This is not exactly unexpected, but I still find it sad that when it comes to performance in functional languages, we must think about the compiler more than we think about the language.