Seeing a lot of delays when cooking LOP nodes on our end (changing any parm triggers a length recook, so we're in the habit of working Manual Update mode now). It's partly due to the complexity of the content we're working with, but I also wanted to get clarity on whether cooking in LOPs took advantage of multithreading?

Like all of Houdini (except for magical SOP for loops and compile blocks) all LOP node cooks are single threaded in the sense that only one node can be cooking at a time. Each node is free to multithread its operation as much as possible though. In LOPs, our ability to do that is limited in many cases by the fact that USD does not support multithreaded editing of USD layers. However once the USD edits are applied, recomposing the stage is multithreaded. Also, primitive pattern evaluations are generally multithreaded. Because of the way hydra works, some parts of the viewport update are single threaded and some are multithreaded (and often it is the viewport update that takes the most time, not the actual LOP node cooking, though LOP cooking can certainly be expensive too).

So short answer: yes and no.

Thanks for the specific answers, Mark. That's great to know!