Hacking for the latest and greatest Xeon CPUs feels a lot like Forth on small machines to me. Each core has only 256KB of fast private memory (L2 cache). Squeezing the working set into that memory means fast execution that is independent of the other cores. That is really worth spending some brain cycles on when optimizing for 36+ core servers.
The more things change, the more they stay the same?
The more things change, the more they stay the same?