The Power7 (2010) already had a very powerful prefetching engine (also programmable) able to detect a multitude of patterns/striding supporting even 12 different streams. Its design, composability and not only automatic detection but its programmability, allowing you to drive its behavior or change according to different computation phases made a fantastic piece of engineering for its time: https://www.redbooks.ibm.com/redbooks/pdfs/sg248079.pdf
Just pointing this out (there are several generations of Power's with improved design) as you mentioned memory controllers. The problem in the blog post is different (aliasing) and I guess that all I can say is that unless some randomness is introduced, every design will have a "worst" access pattern.
Just pointing this out (there are several generations of Power's with improved design) as you mentioned memory controllers. The problem in the blog post is different (aliasing) and I guess that all I can say is that unless some randomness is introduced, every design will have a "worst" access pattern.