People say that they only get a 4x speedup. I've not measured, but it is definitely faster. Not sure why it would be limited to 4x since the number of model evaluations is VERY large.