8800GT G92, 8800GTS G92, the new masters of DirectX 10

Page 2 - Specifications, Promises, Reality

And now, back on the topic, lets examine in detail the technical specifications of G92. The core is based on 65nm process (as for an example Intel’s Core 2 Duo processors, based on Conroe core), contrary to the G80 core, common for all older 8800 video cards, built on the 90nm technology. This makes G92 the most complicated graphic core manufactured by NVIDIA ever, containing 754 millions transistors (64 million more than G80 and 54 million more than R600). This doesn’t automatically mean that 8800 GT and GTS get the “fastest video card” award, 8800 Ultra will keep that honor for some more time because the differences between the mentioned models is not only in the GPU.

As we already said, 8800 GT and GTS are both G92 based – despite the core in the new 8800 GTS is the same as 8800 GT, there are some differences (in favor of the GTS), which were deliberately made by NVIDIA to maintain sales of the more expensive one – 8800 GTS G92. In the following table we present the differences between 8800 GT, 8800 GTS G92, 8800 GTS 320/640MB G80 and 8800GTX / 8800 ULTRA G80.

Lets delve some more into the architectural differences of the video card series based on G80 and the ones based on G92. Already said - G92 has more transistors and its core surface is just about 330mm2. Quite smaller than G80, with its 480mm2 but G92 is still far from the “small chip” definition.

On the other hand, we once more witness the familiar G80 unified architecture. But for G92 we have 8 clusters of 16 – total of 128 stream processors, which at the new, referent GTS model work at 1750MHz, while for 8800 GT they are limited to 112 (one of the blocks is disabled and we have only 7 clusters x 16) and work at 1500MHz.

Unfortunately, the “unlocking freebie” of locked parts of the chip finished with the 6th series of NVIDIA’s GPUs and at least for now there is no discovered way to get fully operational 8800 GTS from a 8800 GT. As a matter of fact, out colleagues at XtremeSystems have already tried to flash a firmware from 8800GTS G92 with modified DeviceID to a 8800GT G92. The video card continued to show its original Device ID which led to speculation that there are several NAND Flash cells, integrated in the GPU, factory programmed and actively stopping wanting overclockers from receiving lots for less…

Block scheme of 8800GT, one of the stream processor clustes is locked

Altogether this architecture resembles G84 and G86 concerning shader clusters more than G80 (we have 4 addressing processors and 8 filtering, and for G92 we have 8 addressing and 8 filtering, which leads to architectural supremacy). Obviously, G92 is far more complicated. But differences don’t sop here. G92 has lost two of its ROPs and now has only 4. But as with G80, every ROP block has 64bit memory channel, as well as L2 cache. Four ROP blocks x 64bit give us 256 bit memory interface. Every ROP can process 4 pixels per clock and if there are four pixel models (RGB color and Z) and the pixels are modeled with one Z component, then every ROP block can process 32 pixels per clock. Other “bonuses” for G92 are the PCI 2.0 interface support, giving double throughput, as well as NVIO integration module, responsible for picture display. Till now it was placed on the board itself, making the PCB even more complicated, due to the G80 inability to include it because of its humongous size. End now in essence…

At the end of the summer when everyone expected anxiously the new NVIDIA and ATI video card generations and more precisely 8800GT and HD3870, the manufacturers in question released rather scarce information to the lusty gamers. AMD bravely showed off 3DMark tests, according to which their flagship beats in performance 8800GT in 3DMark06 and NVIDIA quickly released their card at the end of October so they can make enough sales before the direct competition from ATI, expected on 19th November. But is 3DMakr all that matters and did ATI made a mistake is to be checked and decided now that we can compare the video cards in a single test configuration – without speculation and questions.