Did it improve the final step though? That seems to be single core and I still can't pinpoint how to accelerate it (written a down-thread comment here: https://news.ycombinator.com/item?id=24739796)