The curtain's been dropped on much of Nvidia's upcoming Fermi-based graphics cards, and the five-hundred-and-twelve-core GF100 looks like a behemoth, indeed. A completely overhauled architecture is all about three things: scalability, parallelism, and geometry. Oh, and ripping your eyeballs out.
Tom's Hardware, Anandtech, HotHardware and others go pretty deep on the new architecture, which is now eminently scalable. Here's the overall structure of the GF100, which should give you an idea of the scalability—the GF100 is made up of four graphics processing clusters (GPC), themselves composed of four streaming multiprocessors (which are made up of 32 CUDA cores and texture units) and a raster engine:
To go deeper on architecture, you're better off reading the 10-page reports from any of the sites linked above, but bottom line, Tom's Hardware is predicting something like double the performance of Nvidia's current GTX 285. Anandtech also points out that Nvidia's geometry performance only went 3x between the NV30 engine in the ancient GeForce FX 5800 and current GT200 in the GTX 280, but the Fermi-based GF100 has 8x the geometry performance of the GT200. The endgame being that " it allows them to take the same assets from the same games as AMD and generate something that will look better. With more geometry power, NVIDIA can use tessellation and displacement mapping to generate more complex characters, objects, and scenery than AMD can at the same level of performance."
There is a cost. Even though it's at the 40nm process, those 3 billion transistors are going to run hot, and the GF100 maybe the hottest single-card GPU ever. It's also not going to be cheap. At all. [Tom's Hardware, Anandtech]