The Reference Board
Per usual, the first board we were able to get our hands on was a reference board direct from NVIDIA. Although some add-in card vendors will change the layout and appearance of the board to help differentiate their board from other vendors’, the reference design is typically the blueprint for the vast majority of cards coming to market. Taking a quick glance at the card, we see that NVIDIA has finally given hardcore enthusiasts a dual-DVI option. Those running CRT monitors will not be left out as board vendors will surely make a more conventional model featuring one DVI port as well as the standard D-Sub port.
Looking at the elaborate heatsink assembly on this card, I was expecting the board to have some significant weight. To my surprise, the card is actually incredibly light thanks to the all-aluminum construction of the heatsink. In comparison, both the GeForce FX 5950 Ultra and Radeon 9800XT are much heavier cards due to the extensive use of copper for their cooling solutions. Flipping the card over, we see that all 256MB of memory are now mounted on the front of the card with no modules on the backside whatsoever.
With this much aluminum covering the face of the card, there is little doubt that this board consumes a lot of power and radiates a lot of heat. In similar fashion, we see that the GeForce 6800 Ultra now features two molex power connectors. Not only are you required to use two independent cables from the power supply (no Y-splitters here folks), but the power supply itself must be at least 480W. In all fairness, anyone considering spending $499 on a graphics card should already have a quality power supply. Enthusiasts know the importance of a good power supply with regards to system stability and overclocking success. Should you still be using the generic 350W unit that came with your case, purchase the $399 model and spend the $99 on a top-notch power supply.
Removing the thin cover for the heatsink assembly, we see how dependent each cooling solution is to the other. The fan on the leftmost end of the card is responsible for sucking cool air from the top of the card and pushing it through the fins of both the core and memory heatsinks. Even the power logic gets a dose of cool air for its own tiny heatsink. Disassembling the assembly further, we see that the core’s heatsink is very firmly secured to the board with four spring-tensioned screws. Anchored by the metal plate on the backside of the card, this allows the heatsink to make the best possible contact with the GPU’s surface.
With the memory heatsink assembly removed, we see that the modules used on this board are labeled Samsung K4J55323QF-16. A quick visit to the company website confirms that these are 1.667ns GDDR3 modules rated for 600MHz. Given the default memory speeds of 550MHz, we see that there is ample headroom for overclockers to experiment with. Hopefully, some board vendors will take advantage of the faster 1.25ns modules which are good for an impressive 800MHz.
In an effort to keep the GDDR3 memory modules at a comfortable temperature, the reference board features an interesting heatsink which incorporates a heatpipe. The end of the heatpipe directly above the GPU is filled with a liquid with a low boiling point. As that end heats up, it changes phase to a gas and rises towards the end connected to the aluminum radiator. As the air from the fan blows over the fins of the radiator, the gas is cooled and changes phase back to a liquid dropping to the other end of the heatpipe again. As a result, heat is wicked away from this area faster than the aluminum would otherwise dissipate the heat.
Finally, we have the crown jewel of this reference board. NVIDIA has called upon IBM and their 0.13u process to produce a core comprised of an incredible 225M transistors. This behemoth packs 16 ‘true’ pipelines which ensure that performance will be exceptional.
One interesting feature of this reference board was the intricate ‘band-aid’ we saw on the back of the card. Directly below the fan header, we see a quick fix for some unknown issue. Obviously, this solution will be much more polished on retail boards.
