mnailor wrote:vpo-organist wrote: snip...
If someone says this works and that doesn't work for me, then over time a collection of information forms in the forum that is difficult or impossible to evaluate. Too bad. This is important information from the Hauptwerk customers that should be channeled to get value out of it.
True, anecdotes spread across hundreds of topics are hard to draw any useful conclusions from. But how would you suggest collecting and organizing this?
To me, what is needed is a Hauptwerk capacity planning methodology with a limited number of explicit benchmark tests, a central collection point for users to submit their test results with computer specs and settings, a built-in way to compare test results on similar hardware to identify unreplicable results for human review, auto-updating capacity models to forecast how the benchmarks should do on as yet untested hardware, and a way to compare any sampleset's maximum compute requirements to the benchmarks to help users identify hardware that supports a benchmark that is known to exceed the sampleset's requirements. That's a huge effort.
Of course, my remarks above would be impossible -- that's how the large IT shops I used to work for would handle things, and they had a much more controlled, homogeneous universe of hardware to deal with. Even then, capacity planning and performance modelling was one of the hardest things I did for a couple years...
Maybe more feasibly for Hauptwerk some day:
Samplesets could have an estimate of their minimum required polyphony for each stereo perspective loaded, either by testing or by a crude swag from counting ranks, multiplying by release duration, and multiplying by a constant representing how many notes two hands and two feet can put into play. (Oversimplified but could be approximated with testing.) SP publishes polyphony required for some of their samplesets, so it can be done.
Then if HW had an automated polyphony benchmark (a silent test, please) that would estimate the highest polyphony a computer can handle in the green on the HW CPU meter, using known fixed HW settings, we could contribute those test results along with our computer's specs and audio interface. It would have to be push-button with no special configuration for users who won't rtfm.
Edit: Okay, maybe not a silent test if it matters what the sample buffers sent to the driver contain for testing CPU overhead. I don't know if sending zero amplitude samples reduces driver time in any way.
There'd be plenty of variation in the results because of our OS settings, but even getting a range of achievable polyphony for some computers would let buyers compare hardware to sampleset requirements better than we can today.