As jeff said it targets CPU bottlenecks. One aspect is also that those are techniques that programmers have to apply to their code. So it's not a pure driver optimization that makes anything fast in the driver. The talk as well as others make suggestions to programmers how to organize their render engines to avoid some CPU…
With this low overhead API/driver stuff, it's important to understand that the main target is CPU cycles. So when someone says "it's 15x faster" that's misleading to non-programmers, because what they really mean is "the time your CPU is spending buffering GL calls is now reduced 15-fold". Which is a big deal but not as…