Interesting. One thing I noticed in my test of yesterday was that, when I didn’t have the HALion track there, my Cubase CPU meters were basically doing nothing – i.e. they apparently couldn’t keep up, perhaps due to too much power being claimed by other parts of the program with higher priority (like trying to keep supplying the audio engine with samples)? Thus, at least in that case, I couldn’t make a comparison of behavior with and without the HALion trick. Of course, the Task Manager CPU graphs more or less told the story of how Cubase’s behavior in the two scenarios affected CPU distribution.
My one speculation would be that, when Cubase doesn’t have to worry about the real-time thread, thus assumedly being able to give its all to the ASIO Guard thread (plus stuff like updating the user interface and whatever else needs to go alongside all that) that it somehow slices tasks up into longer segments, and maybe those take too long in extreme cases, thus not being able to get back to the audio interface driver to feed it sample buffers quickly enough. But maybe when the real time thread has to take priority, that forces it to slice things up more finely, reducing the chance of taking too long to feed the audio driver sample buffers. Or maybe the audio driver somehow gets deprioritized in this scenario, making it not be ready when Cubase is???
Whatever is going on, though, it would sure be nice if Steinberg could figure out why this weird (as in contrary to what would be expected) phenomenon occurs then find a way to take advantage of it to make performance smoother without having to add an artificial real time thread just to get better audio streaming results in these high CPU load mixing scenarios.
The thing is, I’m not sure that my i7 5820k at 3.3 GHz, which was released in 2014, would qualify as “modern CPU architecture”. And, since your much newer processor would, the behavior seems to be consistent across CPU generations, not to mention in Windows 10 and Windows 11. It will be interesting to see if the person who mentioned trying it on a Mac also sees a similar thing. If they do, then that would seem to further underscore that it is more a Cubase behavior than a CPU or OS behavior (or alternately, that they should at least be able to rely on a decade plus of CPU and OS behaviors and, potentially, deliver Cubase behavior that takes advantage of this to deliver better audio streaming results).