The GeForceFX has a total of eight pixel pipelines, allowing it to render up to eight pixels per clock-cycle with one texture. When two textures are used (classical multi-texturing), the GeForceFX can process up to four pixels per clock-cycle, and so forth. Per pipe, the FX can apply up to 16 textures in one step. Each pixel pipeline has its own shader unit at its disposal, whose performance (operations & instructions/ clock) is strongly dependent on the code that is used.

A shot from the NVIDA Dawn Demo.

Additionally, NVIDIA has optimized the pipelines for trilinear and anisotropic filtering. Operating at peak performance, the filter-unit is capable of determining the samples required for filtering, which increases performance. NVIDIA has also implemented an adaptive filtering method similar to ATi's in hardware. Based on the textures and the geometry of a scene, the GPU determines at what level a scene should be filtered using either an anisotropic or a trilinear filter, without reducing the image quality. Depending on the perspective and the textures, higher filtering levels are not accompanied by a better image but still cost performance. This "smart" adaptive filter would solve this problem.

Thankfully, the smart filter optimization can be turned on and off through a driver menu, letting the chip revert to a "classical" filtering method. It is gratifying to see NVIDIA leave this decision up to the individual user.

A screeshot from NVIDIA drivers panel (v41.34). The look might change in the final driver version.

The GeForceFX's fillrate is easily superior to the Radeon 9700 PRO's. Eight pixels per pipeline per clock cycle add up to a theoretical fillrate of 4000 Mpixel/s (at a core speed of 500 MHz). This is a good 54% higher than what the Radeon can offer (2600 Mpixel/s).