here we go...
Anand's 640X480 to 1600X1200 benches of the GTS U
analysing this, it appears that the NV16 (what I like to call the GTS Ultra) is still quite a bit bandwidth limited. take a look at this to
bench compare performance with the overclocked NV16 to the normal NV16.
let me go into more detail about the Tile based rendering..
the increase in performance is twofold. first, your rasterizer (the pixel pipes) don't have to go through nearly as many cycles to complete the same scene (they don't have to map a texture to surfaces that aren't seen), which is where that insane Fillrate # comes from.
the second part of the performance increase is a result of the first, becuase you don't have to render surfaces that you cannot see, you don't have to use as many textures (unless they are just repeat textures, used a fair amount on the walls in Quake 3 for example), which means, it doesn't have to load nearly as many textures from the RAM.
this can reduce the Bandwidth requirement by a good 1 gig/second or more (I'm estimating, I really don't have benches), depending on the amount of repeat textures, and the amount of overdraw in that scene.
now, looking at the speed of those benches that the NV16 ran, overclocking the core from 250MHz to 285mhz and the RAM from 230mhz to 250mhz yielded a fair improvement in performance, mostly due to I expect, the RAM overclocking.
I'm beginning to think that nVidia will have to impliment something else really new in the next gen processor (NV30) in order to improve things much, becuase Bandwidth is a major problem now.
HOWEVER by that time, the T&L unit will be a much larger factor, which may have a large leap in performance there (hey we BARELY have any monitors that actually RUN 1600X1200, or higher).
I'm beginning to think that after the NV25, we will see a race for much more revolutionary technology. Voxel acceleration, insane polygon counts for T&L, and more little features that can be applied to textures (little tricks like what comes in DX8, things like EMBM, stuff like that).
we will see a plateau in the fillrate race for a while I think is what I'm trying to say. 300mhz DDR SDRAM will be available to video card manufacturers next year, and after that it will be QDR. that will be pretty good, but even WHEN those come out, the fillrate users will easily be able to fill that kind of pipe.