Implementing O(N ln N) or O(N) gravity on GPUs (which I guess is what is underpinning this) is both much more difficult and it's unclear whether it would actual

Implementing O(N ln N) or O(N) gravity on GPUs (which I guess is what is underpinning this) is both much more difficult and it's unclear whether it would actually run faster for any reasonable number of particles. Don't forget about that constant :)