I'd guess (no real evidence) it could be because a lot (most?) software runs on timescales where 10% isn't that significant. Also, maybe something to do with Amdahl's law, where speeding up one part by 10% rarely speeds up the whole thing by the same amount?