3 ms·

We were using modified lz4 (very much molded over our data) that was actually faster than memcpy of data that we were compressing.