How Big is Fill Rate?

How important is texture fill rate for the next-generation of 3D cards?
Fill-rate is still the most fundamentally important performance metric there is for 3D hardware acceleration. What people don't realize is that many of today's games are not geometry limited whatsoever at interesting resolutions and pixel depths, but rather fill-rate limited. This means that even if you had a Cray Supercomputer sitting in front of your raster engine doing geometry calculations you would not get any performance benefit. Take, for example, a typical TNT2 Ultra card running Quake3 on a Pentium3-600 (Quake3 v1.08, q3testdemo1, vsync off):

…clearly, users want to experience 32bpp color at high resolutions for their games in the future. First look at the 640x480, 16bpp benchmark result which is 84.5 fps. This could very well be a cpu-limited number, it is tough to tell. However, remember that increasing the resolution and increasing the pixel depth does not increase the cpu requirements at all, so looking at the 1024x768 at 32bpp benchmark result (29.8 fps) we realize that this result is completely fillrate limited. The important thing to realize here is that regardless of how fast a geometry engine is, games which are fillrate limited will simply not run any faster when geometry acceleration is enabled. As the GeForce has not increased fill-rate much at all, we expect for many applications the performance to not be substantially above the performance of a TNT2 Ultra product. As we stressed before, we believe the GeForce product to be a very unbalanced product, with good geometry processing capability but perhaps substandard fill-rate capabilites for a next-generation 3D accelerator. So if you want to run your games at 640x480, 16bpp color (a resolution, btw, which certainly has low enough fillrate requirements that the GeForce can show off its geometry prowess) then maybe the GeForce is the right product for you. But we know that gamers want to run in 32bpp color at high resolutions, so the geometry processing power of the GeForce simply does not buy you much…

That being said, 3dfx has always been the fill-rate king, and our next generation product will have fill-rate performance that will simply amaze people. As a result, we are very confident that for the resolutions and pixel depths that gamers really care about, our next generation product will substantially outperform a GeForce product when running Quake3 (and most other applications as well). As I said, if you want to go back to 640x480 resolution at 16bpp color then maybe the GeForce is the right product for you. But we don't think this is the right direction for future 3D acceleration, and we certainly know that gamers demand more. They appear to be delivering a product that's high on marketing checkmarks and low on real-world benefits, but I guess that's their product strategy…