NVIDIA GeForce 8800 GT: The Only Card That Matters

G92: Funky Naming for a G80 Derivative

If we expect the G9x to represent a new architecture supporting the GeForce 9 series, we would be wrong. In spite of the fact that part of the reason we were given for NVIDIA's move away from NVxx code naming was to bring code name and product name closer to parity (G7x is GeForce 7, G8x is GeForce 8), it seems NVIDIA has broken this rule rather early on. Code names are automatically generated, but how we only ended up with three different G8x parts before we hit G9x is certainly a mystery. One that NVIDIA didn't feel like enlightening us on, as it no doubt has to do with unannounced products.

While not a new architecture, the GPU behind the 8800 GT has certainly been massaged quite a bit from the G80. The G92 is fabbed on a 65nm process, and even though it has fewer SPs, less texturing power, and not as many ROPs as the G80, it's made up of more transistors (754M vs. 681M). This is partly due to the fact that G92 integrates the updated video processing engine (VP2), and the display engine that previously resided off chip. Now, all the display logic including TMDS hardware is integrated onto the GPU itself.

In addition to the new features, there have been some enhancements to the architecture that likely added a few million transistors here and there as well. While we were unable to get any really good details, we were told that lossless compression ratios were increased in order to enable better performance at higher resolutions over the lower bandwidth memory bus attached to the G92 on 8800 GT. We also know that the proportion of texture address units to texture filtering units has increased to a 1:1 ratio (similar to the 8600 GTS, but in a context where we can actually expect decent performance). This should also improve memory bandwidth usage and texturing power in general.

Because NVIDIA was touting the addition of hardware double precision IEEE 754 floating point on their workstation hardware coming sometime before the end of the year, we suspected that G92 might include this functionality. It seems, however, that the hardware behind that advancement has been pushed back for some reason. G92 does not support hardware double precision floating point. This is only really useful for workstation and GPU computing applications at the moment, but because NVIDIA design one GPU for both consumer and workstation applications, it will be interesting to see if they do anything at all with double precision on the desktop.

With every generation, we can expect buffers and on chip memory to be tweaked based on experience with the previous iteration of the hardware. This could also have resulted in additional transistors. But regardless of the reason, this GPU packs quite a number of features into a very small area. The integration of these features into one ASIC is possible economically because of the 65nm process: even though there are more transistors, the physical die takes up much less space than the G80.