Can you rephrase your question about the ops perspective? Is there something in particular you want to know about?
HHVM is its own web server (although it supports FastCGI for easier use with existing infrastructure like nginx), and that's how we use it in production. It's hard to compare memory usage, except at scale, where it benefits from not having a whole bunch of interpreters (in different processes) running at the same time and some other benefits by using more appropriate types to store values through type inference.
On a webserver, that goes further since HHVM is single-process/multi-thread, whereas PHP (in its typical setup) is multi-process/single-thread. This cuts HHVM's memory overhead much further.
Yes, PHP can run multi-threaded, but it still has a number of instability issues in that mode.