Upping the L3 from 6 to 8 MB will not have significant performance boost in real world apps. Unless one wants to rotate with approx ~12-16 *N MB dataset and is good at cache thrashing just to demonstrate the importance of cache. Then again, it is also possible to craft a piece of code in which SB's (possibly to be glorified later on by certain person)branch predictor fails 100 %, to demonstrate how stupid the BP actually is.