The G94: Half a G92?
Unlike all the GeForce 8800s introduced since the 8800 GTS 320 MB early in 2007, the GeForce 9600 GT is not based on the famous G92 chip, but on its little brother, the G94. The essential difference is that the G94 only has half the number of stream processors of the G92: 64. This pales in comparison to the 112 a "simple" GeForce 8800 GT 256 MB boasted, you're thinking. That's not untrue, but the shader frequency is higher (compared to the 8800 GT only, since the G92 on the 8800 GTS uses the same frequency), attaining 1625 MHz for these units (650 MHz for the rest of the GPU). But even at that frequency, the raw processing power of this G94 is still a full 38% lower than that of the 8800 GT.
Let's look at the reasoning behind this choice. It's often more interesting for a card maker to attain a given level of processing power by using fewer SPs but a higher frequency, rather than the other way around, because this allows the company to reduce the number of transistors on the chip. This in turn leads to a smaller chip size, thus reducing production costs. Of course, for that principle to work, the required frequency has to be able to be achieved without a severe drop in yields, in which case production costs can end up being higher despite the increase in the number of dies produced per wafer.
Accordingly, the G94 has only 505 million transistors, or 33% fewer than the G92, and it has a surface area we measured at 225 mm2, or 31% smaller, despite the same engraving depth (65 nm). That value is still 15% higher than on the RV670 used in the AMD 3D graphics cards it directly targets, but remember that the latter chip uses 55 nm given its higher number of transistors, and that that necessarily has a financial impact - either at the time of development of the process for the chips, or in terms of yields or production costs.