RE: Expected magnitude for CPU projects
I don't think we can get much out of these statistics, too many variables.
As already mentioned, you don't know how many threads are used or how high the allowed cpu time is nor you know if there's overclocking in use and how high this overclocking is or even underclocking for better performance/power.
As an example I've overclocked some time ago an Intel Core2Quad from 2,4 to 3,6 GHz (=+50%!) on air without any issues (sure, power went way more up than performance in comparison).
Intel's Core-Series had in it's entire life-time not an increase in performance as shown - per core per thread per clock it was not even 2x.
Sure, there are some new instructions (newer SSE, AVX, ...) but I don't think BOINC-Projects use these instructions heavily...
I'm still looking through how to handle outliers in the statistics. The point about overclocking is one of the reasons I chose non-K processor types for intel. On further thought I'm reconsidering whether I should have used the X parts from AMD since they have less overclocking headroom.
The real point of this is to tell you which projects you should look at running beyond the normal advice of picking a project with low team Rac and low total users.
For example, Yoyo has a fairly high Team Rac for a CPU project and a decent number of users. However if you are running a CPU that is from ivy bridge or later (or Zen) it is a good performer across the board. This is probably due to the project not actually having much dependence on the processor frequency.
The tables also show that if you have a high core count CPU you should switch to sourcefinder (can claim more work proportional to # of threads), YAFU, or Yoyo. Which is not something that a new user is going to know.