Fill Rate vs Transform & Lighting
Early reviews of the GeForce 256, which were published during the time NVIDIA's non-disclosure agreement was in effect, caused much debate over the chips' fill rate capabilities. Attention was drawn away from the fact that currently, the GeForce 256 has the highest theoretical fill rate (480 million pixels per second) of any consumer based chip available. Those reviews were also based on early driver revisions which have since been optimized even further. One review actually provided benchmark results using TNT drivers!
Rather than debate on the fill rate versus the transform and lighting issue, let's leave that one to the game developers to decide. Voodoo Extreme continues to maintain a page with comments on this subject from todays leading game developers. Many of the developers see transform and lighting processors as the wave of the future. Mike Dussault of Monolith Productions makes an excellent point on the subject:
The PC bus is the bottleneck we need to circumvent to take PC games to the next level of realism. By using hardware T&L, we can avoid sending the same vertices across the bus over and over. This problem is very similar to the texture memory problem that kills cards.
Hardware T&L allows us to store the vertices on the video card in the same way that we cache textures on the video card. Then we can essentially send a really small command across like "draw a triangle setup using vertices 100 thru 300." By doing this, we've kept the card busy, we haven't maxed out our bus bandwidth, and the processor gets a lot of extra free time to do better collision detection, process more enemies, etc.