I presume this is the case, but were you comparing against PHP with APC or similar in place?
Yep. I double-checked that I hadn't borked the PHP install because I was surprised at the result. I'm not sure what it was doing to make the big difference since the inconsistencies with real PHP behavior were enough to write it off as a drop-in replacement, so I didn't continue experimenting much more, but it seemed promising.
(I didn't end up taking it live because HHVM did not render the page entirely correctly, but the speedup was still pretty nice.)