Good article, but if i had to guess the subsequent L3 cache access after an increment is likely far overshadowed by the overhead of coherence messages, or the cores communicating between each other.