That’s the wrong way to look at it. Improving the performance of a complex piece of software is not something you do in one fell swoop, or even in a dozen smaller steps. It’s a job of compounding many tiny single–digit percentages over years, and of carefully avoiding performance regressions.