I think the technique there is largely covered within the op, but it seems like this may be a somewhat well known technique.

I'm guessing the instruction pipeline gets packed, so 4-5x more than 3 cyles.

If all you care about is throughput, that's what you want

Also, the benchmark of exactly as written there is missing some overhead, as it will always be cached. Memory accesses aren't likely to be an issue here, but I always try to run over a couple dozen or hundred of megs of input in benchmarks to ward against that.