It would be really good if the author actually linked to somewhere where you can reproduce the benchmark. Either way, I'm pretty sure the correct response is "use pypy", not that I can check it