i thought the orders of magnitudes faster benchmarks were around lazy evaluation, and probably wouldn't apply here.

Roc uses perceus, which is a reference counting model that allows for opportunistic in place mutation if it is safe to do so. HVM is more like a fundamentally new evaluation model that is parallel by default. They are both very exciting, but HVM is much more radical and probably needs at least a few more years until it starts to be seriously practical.