My main takeaway from this paper is that the indexing function for cache sets does not necessarily have to be the identity.
It proposed using different ones for each bank and in effect breaks the concept of a "set", but also increases the hit ratio and so claims an effective increase in associativity.