Unified Architecture
Instead of the usual pipeline of dedicated vertex shaders, pixel shaders and ROPs, the new GeForce 8-series of GPUs makes use of large numbers of generic stream processors. There are 128 stream processors in the GeForce 8800 GTX and 96 stream processors in the GeForce 8800 GTS.
Each stream processor is capable of performing vertex, pixel, geometry or even physics operations. The type of operations assigned to each stream processor is dynamically determined by the GPU. This improves the GPU's performance as processing power can be dynamically allocated according to need.
If the application is pixel shader-intensive, then more stream processors will be allocated for pixel operations. Conversely, if the application is more vertex shader-intensive, then more stream processors will be allocated for vertex operations.
The end result is not only a much faster GPU, but also a more efficient one. NVIDIA made it clear that they were gunning for better performance per watt as well as per square millimeter of silicon real estate. Take a look at their NVIDIA's power-performance comparison.
With the GeForce 7900 GTX set as the baseline, it's easy to see that the new GeForce 8800 GTX gives about 52% more performance per watt. Interestingly, the power utilitization of the GeForce 8800 GTX is only slightly (12%) higher than the GeForce 7950 GX2.