Theoretically, the new GPU has 56 rather than 28 texture units, but practice suggests that this approach is not justifiable due to the wide use of anisotropic filtering today. We can recall those quite numerous cases when the GeForce 8600 GTS is slower than the Radeon X1950 Pro: although it seems to have more TMUs, its 16 texture processors turn out to be only 8 actually. This is not so critical with the G92, though. Even if the performance of the texture processors of the GeForce 8800 GT will correspond to the G80’s 28 texture units, the new GPU will still be superior to the GeForce 8800 GTS in this parameter, also because of the higher frequency of the main GPU domain.
Besides the differences in the design of the texture processors, the G92 is almost the same as the G80 on the architecture level. The number of functional subunits is different, though. The new chip has only 4 raster processors instead of 6. Each of them can process 4 pixels with 16 sub-pixel samples per clock cycle, which gives you a total of 16 pixels per clock cycle for color and Z values. If only the Z-buffer is processed, the processing speed is increased to 128 pixels per clock cycle in ordinary mode and to 32 pixels per clock cycle with 4x FSAA. The algorithm of multisampling of transparent textures has been improved. According to Nvidia, it now provides about the same quality as super-sampling but without a great loss of performance. This may mean some changes in the driver, however, because the company used to tout transparency antialiasing as a unique feature of the GeForce 7 series but later turned it on in the GeForce 6, too.