A Closer Look
Power is delivered to the card via a six pin and one eight pin power connector. Both are required to run, it won't work with just one connected.
The GDDR3 memory chips are made by Hynix and carry the model number H5RS5223CFR-N0C. With a cycle time of 1.0 ns, they are specified to run at 1000 MHz.
Eight of these Renesas R2J20651NP DrMos MOSFET chips are located on the board, four for each GPU. Each one incorporates a high-side FET, low-side FET and driver in a single package - essentially one PWM phase. Such a high-density design helps to reduce the PCB area consumption and also reduces the complexity of sourcing parts.
Instead of the well-known Volterra voltage controllers that allow software voltage control, NVIDIA has chosen to go with the ADP3193A voltage controller this time. Unfortunately it doesn't support I2C which means software voltage control is not possible. The datasheet is available online, so voltmodders can easily have a go at hardmodding their cards.
NVIDIA has separated the display output logic from the GPU on the latest chips. So in order to drive the two DVI outputs a single NVIO chip is required. Why there is a second chip, I don't know, it shouldn't be needed to drive two outputs.
NVIDIA is using their own NF200 PCI-Express bridge chip to interconnect both GPUs. AMD uses a chip from another manufacturer for their HD 4870 X2. NVIDIA's chip offers some additional features which are supposed to help with SLI performance, we have also seen several motherboard designs on which the NF200 is used.
Here you can see the two GT200b GPUs that power the GTX 295. Each which is made in a 55 nm process at TSMC Taiwan with 1.4 billion transistors.