I recently spent some time optimizing a small Julia program I wrote that generates a lookup table of brainfuck constants. Because it only needs to run once, I originally didn't care about performance when I originally wrote it (and the optimization was mostly for fun).

I achieved an ~100x improvement by adding types, using static arrays and memoization. In the end, the performance was mostly limited by primitive math operations, I tried using multiple threads, but any synchronization destroyed the performance.

However, the most impressive thing was the ability of Julia to scale from dynamically typed scripting language to almost a compiled language with minimal changes to the code.