Isn't false sharing (as in all your addresses hashing onto the same cache lines) still an issue for power-of-two sizes? You'd have to mess with padding to figure out what's fastest for each chip regardless of core count.