I read an interesting comment that the quantum efficiency of the current generation of sensors is already ~50%. If noise and quantum efficiency are linear, then doubling quantum efficieny would give you 100%, and we'd get 1 stop better noise reduction. The author of the post I read (jrista) was making the point that even 1/3 improvement in high ISO noise at this point would be very impressive. Low ISO noise is a different beast, mostly cause by read noise and Canon's Analog Digital converters. That could be improved.

Hopefuly I am representing the post correctly.