I know this probably isn't intended for performance, but it would be fun to run this in cuNumeric [1] and see how it scales.

[1]: https://github.com/nv-legate/cunumeric