icrontic.com Writes

NVIDIA’s Director, Technical Marketing Tom Petersen today presented a video on YouTube that benchmarks the upcoming GeForce GTX 480 against the Radeon HD 5870 in Unigen Heaven, a DirectX 11 benchmark famed for heavy use of tessellation.

After giving a brief overview of the core hardware (Intel X58, Core i7 960 and a GeForce GTX 480), Petersen puts the 480 to work in the benchmark routine built into Unigine Heaven. Graphs presented in the video suggest that the GTX 480 is dramatically faster than the 5870, but it goes without saying that corporate benchmarks should be taken with a lump of salt.

It must also be noted that the Fermi architecture behind the GTX 480 is particularly suited to tessellation work, which the Heaven benchmark uses in abundance. This is because Fermi is capable of roping any number of its 16 stream multiprocessors (SMs) into tessellation work, which can lead to dramatic performance advantages when tessellation is the focus. This design differs from the Radeon HD 5800 series which has a dedicated tessellator unit, or a fixed-size tessellation engine that offers a specific and consistent tessellation performance.

The “issue” with NVIDIA’s design, however, is that those SMs are likely to be busy with other tasks when it comes time to render a game. This means that tessellation-focused benchmarks like Heaven can return sky-high figures for Fermi-based cards, as the GPU is not spending the majority of its time doing the mundane rendering work it would do in a game.

Secondarily, we know from NVIDIA’s marketing materials that the company placed a particularly high focus on floating point capabilities. Float horsepower does very for little games, but it is particularly potent for tessellation, a GPU task that is as math-y as they come.

So, if you’ve ever wondered why NVIDIA has been preternaturally obsessed with the Heaven benchmark, the short answer is simple: it uniquely leverages the Fermi architecture in a way games would not to return benchmark results that shame ATI scores.