NVIDIA decided to go with a Unified Shader Architecture with the G80 chip. The benefits of Unified versus discrete Pixel and Vertex Shader units is that the unified shader can be allocated to either Pixel, Vertex, Geometry or Physics shaders as needed by the application. In a scene where more Vertex Shading is required, the G80 can allocate more of its Stream Processors to Vertex Shading.

The 8800GTS has 96 individual Stream processors. Each processor is capable of being used for pixel, vertex, geometry or physics operations. The classical graphics card had many pipeline stages for each of the major stages. The pixel shader alone on the GeForce 7 required over 200 pipeline stages. The G80 architecture streamlines the number of stages because of the nature of the Unified Shader, meaning that the data will move from the shader core for processing to the top of the shader core to continue processing until all shader operations are done and the fragment is sent to the ROP to write to the memory. Each Stream Processor is fully decoupled, scalar can dual-issue a MUL and a MAD (Multiply and Multiply Add) and supports IEEE754 floating-point precision.

Each Stream Processor is clocked at 1.2 GHz. You heard that right, 1.2 GHz. The GeForce 8800GTS has three different clock domains, the core clock, the memory clock and the shader clock. The 8800GTS does approximately 390 Gigaflops of shader power, but is much more efficient with its floating point operations than a traditional architecture with instructional issue limitations like the x1950 which can only issue 3+1 instructions in a clock.

In gaming situations, there generally has been more need for pixels than vertices, hence the classical graphics card came with a higher number of pixel shader units than vertex shader units. There are situations where the application needs more vertices than pixels, and the graphics card is limited to the vertex speed. Take for example a GeForce 7800GTX with 24 Pixel Shader units and 8 Vertex Shader Units. There are instances where the card will only go as fast as the vertices are fed.