The 7800 GTX GPU (formerly known as 'G70') is based quite closely on the 6800 Ultra graphics processor. That said, there are several significant improvements and tweaks which should add up to a large overall performance boost.

Let's get the nuts and bolts out of the way first. The key architectural differences between the G70 and the 6800Ultra start with the size and manufacture of the GPU. The G70 is based on TSMC's 110nm process and incorporates a mind-boggling 302 million transistors. To put this in perspective, an Athlon 64 FX-55 has only 105.2 million! The core is 18.5 x 18.5mm in size, quite large considering it's built on a 0.11 micron process.

The reference 7800 GTX runs at 430MHz to start with, and uses 256MB of GDDR3 memory clocked at 1.2GHz. This provides the reference card with a total bandwidth of 38.4GB/s.

Internally, the 7800 GTX has 8 vertex shader units to the 6800's 6, and a total of 24 pixel pipelines (six groups of four quad pixel operators) to the older GPU's 16. The 7800 GTX will boast a fill rate of 10.3 Gigatexels, considerably more than the 6800 Ultra's 6.4 Gigatexels. The 16 ROPs (Render OPerators) are the same amount found in the 6800 Ultra, meaning that it can't actually render any more pixel data to memory than could the NV40. According to nVidia, this wasn't a bottleneck preventing higher graphics performance.

Smoke and mirrors = eye candy

Up to this point, everything is as expected. Newer GPU equals more pixels per clock plus more bandwidth equals faster. Duh. Hold on there though... Things might not be quite as they seem. The 7800GTX actually marks a significant shift in emphasis by the nVidia design team towards effects and shader tricks as opposed to raw rendering power.

Yes, the GPU is faster than its predecessors (and apparently its current ATI competitors) and will get faster still, but the most significant architectural increase is inside the pixel pipelines. the 8 additional pixel shader pipeline do massively increase fill rate, but all the pipelines benefit from a major internal improvement: enhanced MADD (Multiplication-ADDition) capabilities.

Each pixel pipeline has two ALUs (Arithmetic Logic Units) which perform adding and multiplying operations on each pixel to apply various graphical effects. In the 6800 GPU, only one of these ALUs could perform true MADD (both multiplying and adding) operations which are important to many 3D effects. With the 7800 GTX, both ALUs are capable of 2 MADD operations each, vastly increasing the power of the GPU in certain operations which nVidia feels will be necessary for gaming in the future.

This improvement to the pixel pipelines does not increase the overall fill rate, but it does affect what the GPU can do with each pixel on each clock cycle. The additional MADD operations will allow for much faster rendering of complex transform and lighting effects and image maps, which sheer fill rate alone cannot compensate for.