That's the interesting thing; they don't seem to have.

If they'd gone for ref-counting benchmarks, that would have shown something more like 5-10x; it's really good at that. The graphs they showed seem to be reflective of performance in general-purpose benchmarks (SpecINT et al).