Benchmarks are currently not based on published code and not reproducible, so once time allows I'll figure this out and mark as closed.