Let's get the hard data out of the way first: 480 CUDA cores, 700 MHz graphics and 1,401MHz processor clock speeds, plus 1.5GB of onboard GDDR5 memory running at 1,848MHz (for a 3.7GHz effective data rate). Those are the specs upon which Fermi is built, and those are the numbers that will seek to justify a $499 price tag and a spectacular 250W TDP. We attended a presentation by NVIDIA this afternoon, where the above GTX 480 and its lite version, the GTX 470, were detailed. The latter card will come with a humbler 1.2GB of memory plus 607MHz, 1,215MHz and 1,674MHz clocks, while dinging your wallet for $349 and straining your case's cooling with 215W of hotness.
NVIDIA's first DirectX 11 parts are betting big on tessellation becoming the way games are rendered in the future, with the entire architecture being geared toward taking duties off the CPU and freeing up its cycles to deliver performance improvements elsewhere. This is perhaps no better evidenced than by the fact that both GTX models scored fewer 3DMarks than the Radeon HD 5870 and HD 5850 that they're competing against, but managed to deliver higher frame rates than their respective competitors in in-game benchmarks from NVIDIA. The final bit of major news here relates to SLI scaling, which is frankly remarkable. NVIDIA claims a consistent 90 percent performance improvement (over a single card) when running GTX 480s in tandem, which is as efficient as any multi-GPU setup we've yet seen.
