We held the unified shader architecture first implemented in G80 in high regard due to its all-around superiority to the separate vertex/pixel shader designs used in every (modern) GPU prior. The GeForce 8800GTX has 128 of these unified shaders. The GeForce 8800GTS has 96. The GeForce 8800GT is to have 112. This is a pretty straightforward and basically linear comparison to make: the 8800GT has 16 more shaders than the 8800GTS and 16 less shaders than the 8800GTX. This suggests that, all else equal, the card D8P-based card will perform somewhere in between its siblings. Interestingly, there are actually 128 shaders on the D8P core, but 16 of them have been disabled. When producing the actual GPU die, many of the units on the silicon wafer turn out to be partially defective. This is true in every brand of chip making, and is handled differently by different companies. IBM ran into this problem when producing Cell processors for Sony’s Playstation 3. The end result was the announcement that there would actually be fewer SPE’s on production PS3’s (a.k.a change the spec). AMD is reportedly running into problems with their upcoming Phenom processors. As a result they will bring 3-core processors to the market, which are effectively quad-core processors with one core disabled. This practice ends up reducing the manufacturing costs of the chip, and decreases the amount of chips that are thrown in the garbage. For the consumer, lower shelf prices are the result of creative rebranding and remarketing of less than perfect parts. NVIDIA’s approach to increasing usable yield is very similar: D8P cores with 112 shaders used on the 8800GT are just 128 shader cores with 16 disabled. More D8P cores per wafer can be used if 16 defective shaders can be disabled. There is nothing wrong with yield-improvement techniques, and certainly if it means lower prices for the consumer we are all for them. Advantage on shaders: 8800GTX.


Core, memory, and shader clockspeeds are all refreshed on the 8800GT. One benefit of a die shrink (65nm from 90nm) in any core is the ability to increase clock speeds with little to no additional heat overhead. We saw this when Intel moved its Pentium D line to 65nm, and we are about to see it again from them as they roll out Penryn, the long awaited Conroe refresh. The reference D8P board we sneaky snaked from NVIDIA is has a core clock of 600MHz, a memory clock of 900MHz (1800MHz effective), and a shader clock of 1.5GHz. All of these numbers are equal or beyond what stock 8800GTX numbers, 575, 900, and 1.35, respectively. Not only this, but several of NVIDIA’s add-in board partners will be shipping the 8800GT with factory overclocks, up to 660MHz on the core. This is a pretty large factory overclock, and is no doubt the result of a tremendous amount of headroom on the D8P core. We did not run any tests when our cards were overclocked, but we had two 8800GT’s running in SLI pretty comfortably at 690/1900. Advantage: 8800GT.


The last really big changes from G80 to D8P are the amount of memory and the memory bus width. On the 8800GTX, there is 768MB (512+256) of GDDR3 memory supplying the core through a 384-bit (256+128) bus. On the 8800GT, there is 512MB of memory (with a 256MB supposedly soon to follow) supplying the core through a 256-bit bus. The amounts here are basically the same if you disregard that weird extra stuff on the G80. I don’t know if anyone ever really figured out what the hell that was for anyway. However, numbers being numbers, there is a clear advantage here to the G80 based 8800GTX. The increased memory amount should lead to increased performance with games that use very high texture detail with very large environments (expect Call of Juarez and Oblivion numbers to benefit). On a side note, the 640MB of the 8800GTS also has more memory than the 8800GT, so we might expect that to make up for some of the ground lost as a result of fewer shaders and slower clock speeds. Advantage: 8800GTX.


So really what we have here is a battle between the new GeForce 8800GT and the one year old GeForce 8800GTX. Each card has something going for it that in theory would put it a hair above the other, so it should be interesting to see which card excels in which areas.

In terms of physical features, the 8800GT is a pretty awesome video card. If the specs are any indication, it would seem that we have what basically amounts to almost an 8800GTX on the same size PCB as the 8800GTS (9.5” vs. 11”), and is cooled using a single slot fansink design (as opposed to the 2 slot design used on the GTX and GTS). The actual fansink on the 8800GT is very nice looking. It is sleek and extends the length of the card. It appears to have a copper heatpipe integrated into it somehow, and the reference design sports a chromed NVIDIA logo type thing on the side. The annoyingly loud fan (under load) expels air through some vents at the top of the card. Said air is directed towards this vent using a neat heatsink fin design that resembles the spiral of a conch. The air seems to leave the card in a manner that would allow it to shoot right out of the back of the case if it has a vent adjacent to the add-in board slots. The card has two Dual Link DVI ports and one S-Video port in the back for connectivity, and is powered by just one 6-pin PCI-Express connector supplementing the PCI-Express slot. Specs for the card indicate that it requires a 400W power supply, or a power supply with 26A on the +12V rail. These numbers seem pretty damn high considering the maximum power draw for the board is a mere 105 Watts.