> if performance is more important then easy portability, you need to implement driver specific code paths anyway
I think that's fine, because as you said, it encourages GPU vendors to innovate. Plus, it probably affords more of an opportunity to squeeze every last drop of performance out of each card.
On the other hand, I'd start to get upset if we had to write card-specific code just to do the absolute basics. (Which isn't the case right now.) There are tons of apps that need only a tiny fraction of the performance a GPU offers. For those apps, it would be insane having to maintain multiple codebases just to, say, draw a box with a texture and Lambert reflectance.
Other then that I don't even think of OpenGL as a single standard anymore. It's more like "Nvidia GL" "AMD GL", "Intel GL", "Apple GL", etc... there is a core set of functionality which works across all implementations (and which could be cleaner), but if performance is more important then easy portability, you need to implement driver specific code paths anyway. Whether this is good or bad I haven't yet made up my mind completely. At least on GL you have an environment where GPU vendors can experiment and compete through extensions.