#1 thing we figured out is that 'raw performance' and 'perceived performance' are two very different things - we were selling raw performance, customers were buying perceived performance - perceived performance is sort of an S-shaped curve, on the left there's a flat area where it's too slow, increase it a bit, it's still to slow, in the middle there's an area where performance changes are very noticeable, and on the right there's another flat area where it's fast enough.
Getting to that middle bit first was where we made our money, once we hit the rightmost bit we were mostly waiting for the competition to catch up.
Anyway the real point I'm trying to make here is that what you're aiming for is that 'fast enough' spot, not 'fastest' - for everything, it's worth spending time benchmarking to figure out what is slow (spending the time watching someone else use it is time well spent)