Well, it depends. Running the exact same operations through shows the difference to be negligible (the text paths were really optimized). Of course, there are some particularly bad (and rare) paths in the text protocol that don't have exact analogs in the binary protocol which helps to keep it cleaner.
However, clients can safely perform certain types of optimizations now that should lead to a lot of improvement.
One thing to know is that we get really high operation rates through the current version (higher than the really high operation rates we got through 1.2.8), but they're all generally faster than you can process them.
That is, we can generally saturate network interfaces before any other part of your servers.
This is to the point where some users replicate not for fault tolerance (because that's rather silly), but to increase read capacity.
