The conversion to a .25 micron process, and probably some architectural refinements, have enabled default clock speeds that are 40% to 66% higher than the first iteration of the TNT. In combination with aggressive memory settings that keep the 3D engine fully occupied, TNT2 is a superb performer. There is no excuse not to play games at a resolution of 1024x768 in 32-bit color. Nevertheless, proper support for trilinear filtering is a fairly non-trivial issue that needs to be addressed.
Trilinear filtering is an important image quality enhancing technique. There are some that would have you believe that trilinear filtering does not make a perceptible improvement over bilinear. The previous three 'Unreal' screenshots captured on a TNT2 (bilinear and 'trilinear dithering') and Savage4 ('trilinear approximation') are good starting points to make an informed decision.
Equally important is the motivation behind 'trilinear dithering'. It is not acceptable on the part of the device driver to initiate 'trilinear dithering' when in fact the application is specifically requesting trilinear filter. The correct method of applying true trilinear on two textures involves four bilinear fetches. Since this is a two-pass technique, with the exception of the Savage4, there is a performance penalty. Is the TNT incapable of applying trilinear on two textures in multiple passes? If that is the case, 'trilinear' should not appear as an option for the TNT in multi-texturing games. One can only hope that this is an issue that will be resolved by modifying driver behaviour.
The other issue relates to emboss bump mapping, which looks set to be a major feature in upcoming games. As Mr Derek Perez has pointed out, the coding example published by NVIDIA used white specular light by default since blending of the interpolated diffuse color and texture color was omitted. The question is whether this additional blending will require an additional pass, with its obligatory performance penalty.
By Kert Chian