I didn't say it was the only reason, but it is a big reason. Even if you use a depth map to exit early on fragments you are still going through all your geomet

I didn't say it was the only reason, but it is a big reason. Even if you use a depth map to exit early on fragments you are still going through all your geometry twice, once to get your depth map and again to do all your shading. The history of renderman (prman) went in a similar same way, though not through using actual g-buffers. Originally it would shade and then hide because of memory constraints and displacement of micropolygons happening in the surface shader and not in a separate displacement shader. Eventually the architecture shifted to try to do as much hiding as possible and split out displacement shaders into a separate type. Another thing that can be done with g-buffers is using mip-mapping and filtering of the g-buffer to do some sampling or effects on lower resolutions. Pixels can also be combined into distributions instead of just averages, though I'm not sure how much this is done currently.