GeForce 7800 GTX in Detail

Getting closer to practice, it’s time to take our sample of the GeForce 7800 GTX in our hands. This graphics card resembles the GeForce 6800 GT at first:

Both cards use a compact single-slot cooling system, and the component layout of the GeForce 7800 GTX resembles NVIDIA’s earlier products, too. Changes are most visible in the rear part of the PCB where the voltage regulators and other power elements reside. It’s not that the power circuit has become simple, but there are fewer electrolytic capacitors and the power elements are placed in three rows rather than in a single line as before. They form a small rectangular now, covered with a thin-ribbed aluminum heatsink. These changes in the power circuit layout have made the PCB of the GeForce 7800 GTX longer, so the new device is obviously the longest graphics card today.

There is an aluminum plate on the reverse side of the PCB, besides the usual bracket for fastening the cooler. The plate is not just a decoration as one might have thought. The thing is the PCB of the GeForce 7800 GTX is intended for 512 megabytes of GDDR3 memory, and there are 16 places intended for memory chips, 8 places on each side of the PCB. But the standard amount of memory on a GeForce 7800 GTX card is 256 megabytes. They could have put eight 256Mbit chips on the face side of the PCB, but NVIDIA preferred to install four chips on either side. So, the above-mentioned plate is a heat-spreader for the four reverse-side GDDR3 chips. Its efficiency may be not very high, but GDDR3 memory features a low level of heat dissipation, and there are only four such chips to be cooled there.

The graphics card uses 1.6ns memory from Samsung rated for 600 (1200) MHz frequency. This is exactly the frequency the memory chips are clocked at on this card. Note that starting from the GeForce FX 5900 NVIDIA places memory chips in such a way as to make the pathways that connect them to the GPU as short as possible. This helps to ensure stable operation at high frequencies.

We dismantled the cooling system to access the graphics processor:

As you see, the die area of the G70 is much larger than the NV40 notwithstanding the thinner tech process (0.11 micron against 0.13 micron). No wonder as they have added 80 million transistors more into the new chip. The surface of the G70 is not mirror-like like the surface of the NV40, but kind of matte, probably due to difference in the tech processes: the NV40/45 is manufactured at IBM’s East Fishkill facilities, while the G70 at TSMC’s fabs. The shape of the dies is different, too. The G70 is a square whereas the NV40 is a rectangular. There is no separate HSI die here ?the G70 natively supports PCI Express. Our sample was manufactured during the 16-th week of the current year, i.e. somewhere at the end of April, and this indicates that NVIDIA doesn’t have problems with manufacturing the new chip. The symbols “A2?denote a second revision of the chip, and this too makes us hope that the supply will be sufficient.

The cooling system deserves to be discussed separately. It is a variation of the GeForce 6800 GT cooler, but a seriously improved one. The blower is driving air through two aluminum heatsinks joined with a U-shaped heat pipe and is also cooling the heatsink on the power regulators of the card. The heat pipe doesn’t only transfer heat from one heatsink to the other, but also takes heat off the needle section that touches the memory chips. The whole arrangement is covered in a plastic casing, although NVIDIA used to employ a metal casing earlier.

Curiously enough, it is perfectly visible in reference snapshots from NVIDIA that the first or main heatsink that takes heat off the graphics core is made of copper, while on our sample it was made of aluminum. Why? We suppose that copper coolers will be mounted on advanced versions of G70-based graphics cards, with 512MB of memory and clocked at higher frequencies, while the current version of the GeForce 7800 GTX is quite satisfied with the aluminum cooler.