Last November NVIDIA stunned the 3D graphics world with the debut of their GeForce 8800 family of graphics cards. Initially launching with two cards, the GeForce 8800 GTX and the GeForce 8800 GTS, the 8800 family delivered revolutionary performance with the added benefit of DirectX 10 support.
The G80 GPU housed inside these cards is built like no other NVIDIA GPU before it; this time around NVIDIA has started with a blank piece of paper.
G80’s revolutionary architecture
Instead of a vector-based GPU like G70, for G80 NVIDIA has incorporated a scalar architecture. NVIDIA’s engineers found that scalar operations were being used more frequently and that these scalar computations were difficult to compile and schedule efficiently on a more traditional GPU like G70. For G80 NVIDIA also broke up the traditional pixel and vertex shaders as we’d previously known them, opting instead to integrate their so-called stream processors.
NVIDIA’s stream processors are fully programmable and are equally at home working on pixel, vertex, or geometry shading. They can even be used to handle physics processing when Havok FX games begin shipping later this year. This unified design can lead to substantial performance and efficiency gains over previous architectures. To deliver breakthrough shading horsepower NVIDIA has packed in dozens of stream processors inside G80, in the case of GeForce 8800 GTX, NVIDIA integrates 128 stream processors, while GeForce 8800 GTS boasts 96 stream processors. The stream processors run at their own clock speed that’s independent of the rest of the graphics core. In the GeForce 8800 GTX for instance, the stream processors run at 1.35GHz, while the rest of the GPU runs at 575MHz. To keep the GPU fed with data, the GeForce 8800 GTX also sports a wide 384-bit memory interface, providing up to 86.4GB/sec of peak memory bandwidth. That’s just over 35GB/sec more memory bandwidth than the GeForce 7900 GTX. Keep in mind that NVIDIA’s GeForce 7600 GT tops out at just 22.4GB/sec.
All this new circuitry comes at a huge cost though. G80 is easily the most complex graphics processor that’s been released to date, consisting of approximately 681 million transistors. That’s over twice the number of transistors as NVIDIA’s previous flagship, G71, which contained around 278 million transistors, and both GPUs are built on the same 90-nm manufacturing process. Just for a little perspective, even Intel’s quad-core Core 2 Extreme QX6700 CPU weighs in a little slimmer at 582 million transistors and it’s built at 65-nm. What does this mean for consumers? In short, GeForce 8800 chips are pretty costly for NVIDIA to make, particularly in comparison to their previous generation GeForce 7900 GTX cards. As a result, street prices for GeForce 8800 GTS cards start around $400 ($350 after mail-in rebate at some e-tailers) and quickly go up from there, while GeForce 8800 GTX cards typically sell for $550+ online.
In other words up to this point GeForce 8800 cards have been prohibitively expensive. In order to reach critical mass, a cheaper GeForce 8800 was needed. Today that card has arrived in the form of the GeForce 8800 GTS 320MB.