I'll admit I read the architecture document and am now a little disappointed.
I'm working with a couple people in my group to do a pretty significant reinstall of our SW installation, and in our findings, NCM can be a beast when it comes to system utilization. The separate system at least allowed us to keep any of our NCM tasks from impacting the main poller. However, I guess we can shove all the NPM polling tasks onto a separate polling engine for that purpose.
Ah, well.