Nvidia's earlier 7950 GX2
Unlike AMD and the legendary 3dfx, NVIDIA is faithful to the trend it started with the GeForce 7950 GX2 and uses two PCBs on the GeForce 9800 GX2, which cannot be seen in the photos due to the cooling system cover. Dual PCBs have their advantages, the first being easier wiring of the memory chips. Numbers favor NVIDIA, as the manufacturer was able to maintain a 1 GHz clock (970 MHz for the 8800 GTS 512 MB) when AMD had to content itself with 900 MHz on its Radeon HD 3870 X2 build using one PCB. That's even lower when compared to the 1125 MHz of a single Radeon 9870.
Unlike the 7950 GX2, the 9800 GX2's two PCBs face each other and transmit heat to the same heat sink, which is cooled by a single fan. If this approach increases cooling constraints (both GPUs are close and a single heat sink must dissipate the accumulated heat), there are no significant temperature gaps between the GPUs (a problem partially solved on the 3870 X2 thanks to the heat sinks). The small amount of heat transmitted by the GPU to the PCB is also spread more efficiently since there are two PCBs.
Unfortunately, NVIDIA was unable to squeeze all of the components onto a card design that was shorter in length. The 9800 G2X's length of 26.7 cm is typical of a high end card (8800 GTX, Ultra and 3870 X2). Likewise, two power connectors (6 pins + 8 pins) are still necessary and finally, only a fraction of the heat is sent outside the case. Most of the heat is sent to the top of the card and is therefore recycled inside the case, since the vent on the brace is too small due to the number of outputs.