This was interesting for me: it's some benchmark data collected over all versions of this program starting from version 0.4 and up to the compiler in version 0.7. Most of the effort in tuning the program has paid off. The compilation occasionally loses vs. interpretation in very short programs or those that don't loop much (you can see this effect in dynamic and dderiv.

Otherwise, in die-hard computational scenarios like ack and puzzle, the compiler is seen at its best.