Tessellation: Unigine Gives Us Synthetic Numbers
For the past nine months, we’ve been running Unigine’s Heaven benchmark as a measure of tessellation performance. I’ve never really been happy with that. After all, Heaven is an engine demo, and it’s a little disingenuous to make judgment calls based on what largely boils down to a synthetic measurement.
Fortunately, HAWX 2 emerged right after the GeForce GTX 580 launch. The game is heavy on tessellation, so Nvidia was naturally all about it. But I made a conscious decision not to roll the HAWX 2 benchmark into testing prior to launch, as AMD suggested that it might be improving its performance picture through a patch or driver that’d dial down the game’s use of tessellation. Neither materialized, though, and now we’re facing a retail title that people are actually playing. Suddenly, its performance becomes much more relevant.
The exciting thing about HAWX 2 is that it’s tessellation for the sake of improving image quality. We’ve already seen tessellation for the sake of improving performance in Civilization V, along with tessellation sans any real gratification in DiRT 2 and Aliens Vs. Predator. So this was a new one for us.
First, let’s create a baseline by benchmarking the latest crop of cards in Unigine’s Heaven test.
Clearly, Nvidia’s cards handle the Extreme tessellation setting better than AMD’s, but using Normal tessellation, AMD still fares pretty well. Claims that the company improved tessellation performance in its Radeon HD 6800-series seem fairly well founded by the 6850 CrossFire’s showing against the Radeon HD 5970.
Also interesting is that the GeForce GTX 570 is faster than the 480 in all three tests. This makes sense, since we're talking about the same number of PolyMorph engines in play, but higher clocks on the GTX 570.
A look at scaling tells the whole tale, though. The AMD cards are brutalized by the use of Extreme tessellation. They’re less affected by Unigine’s Normal setting, but it’s still an overall loss to Nvidia.
How about the fact that, regardless of whether the card features 16, 15, or 14 PolyMorph engines enabled, all of Nvidia’s tested products seem to scale pretty evenly? It certainly appears that something other than geometry processing power is holding our GTX 500- and 400-series cards back here—or at least, the PolyMorph engines aren’t scaling as well as Nvidia might have otherwise implied.