The vertex cache of the TNT is probably the largest among the graphic processors, being able to accommodate 16 vertices. Perhaps the sheer size of the cache helps TNT2 achieve a peak triangle throughput of 9 million triangles per second. The increase from 6 million triangles per second in the original TNT is mostly a function of clockspeed. The 'advanced floating point setup' forms part of the 'TwiN Texel pipeline'. As 'advanced' as the floating point setup may be, one of the main bottlenecks in graphics involves transferring triangle vertex data across the AGP bus. In a presentation by NVIDIA, it is estimated that transferring 2 million triangles in one second requires a system memory bandwidth of 0.6 to 1 gigabyte per second; a value that will completely saturate an AGP 2x bus.

Thus, comparing triangle throughput among different graphic processors is quite pointless if the central processing unit doesn't deliver. Perhaps when 'hardware transform and lighting' becomes a reality, the issue of triangle throughput may be of relevance. Besides 'floating point setup', the 'TwiN Texel' pipeline contains 'pixel processor' as well as 'lighting & special effects processor'. The 'lighting and special effects processor' probably refers to the texturing pipeline. Peculiar to the TNT series, the pixel processor precedes the texturing processor in the pipeline. By possessing two pixel processors, the TNT series can theoretically deliver two pixels in one clock cycle. However, the usefulness of this is questionable as games use multiple textures extensively. It is doubtful if the graphics processor can switch rendering states on-the-fly to achieve maximum output.