come up with some NEW ideas that havent been thought of yet and implement them in Software and wait for the hardware to catch up
How about profiling registers for optimization? (like the pentium)
If you look at games, most time is spend in the rendering pipeline…
An effective and universal texture layer combination api is required …
Supporting about 8-element arbitral polynomials for each set of layers, and some sources having a constant coefficient and two value sources, one of each is inverted (1/x) … this will allow to make effects possible … specular bump mapping with reflections affected by perturbed normals will be a kids game. Any suggestions ? Contact me at xm@x-infinity.ru - wish to change the OpenGL world 
one cool extension and a good piece of silicon will do the job.