Or you just go with "arrays of 64 byte multiple atomic structures" and you get parallelism and cache!