I recently experimenting with this on PBCS.
Others experience my be different - please chip in if you have - but I noticed virtually no performance difference when comparing with a baseline aggregation that originally had intelligent calc off, with a aggregation that was taking ~2 minutes to run.
This was when experimenting against 3 different scenarios; -
1. Previously processed, so you would expect the dirty only process to have most effect
2. Only a single combination of data dirty, so again you would expect a considerable improvement processing dirty only
3. After a very large scale change when you would expect minimal difference between the two aggregations
My only problem is with PBCS tuning benchmarks is sometimes my aggregations are massively variable with no user load on the system (up to a factor of 3 times longer) which I am guessing is down to share of physical server cpu on the Oracle Cloud server farm...
Mmmm - not a glowing endorsement, anyone else have a genuinely good result using the settings I describe to use intelligent calculation in a less than full cube aggregation?
I also have massively variable aggregation times with the same script and no database load to explain it...
In the end I followed your lead and ran a few tests, and sadly reached the same conclusion.
Couldn't find anything in the log files that made it obvious that intelligent calc was not processing clean data, so would be interested if anyone can confirm that, but for now closing this question as no other useful input is incoming and from my own tests it does not seem to help.