Tessellation

As you already know, high tessellation performance is one of the key architectural advantages of the Fermi family, and the new GeForce GTX 580 in particular. Tessellation is perhaps the most important feature of DirectX 11. Parallel geometry processing used in all NVIDIA's modern solutions is very efficient in tessellation with high levels of primitive decomposition, when the numbers of triangles are very high. This is no surprise, because primitives are simultaneously handled by 16 PolyMorph Engines, unlike a single (albeit more powerful) unit of competing solutions.

We have already mentioned that AMD criticizes the use of very small triangles in certain benchmarks (Heaven, HAWX 2), considering tessellation of that level inefficient. This is true, but only for previous architectures featuring traditional geometry pipeline (capable of processing one triangle per clock).

While in case of the Fermi architecture, especially top-end processors like GF110, such tessellation is handled quite efficiently, which is also indicated by performance benchmarks. Complaints about triangles being too small seem strange. Just look where 3D graphics is headed for these days, those rendered cartoons, for example. Numbers of triangles are much, much higher in those.

Obviously, such quality isn't reachable for computer games yet, but the industry is clearly headed that way. Chances are that GPUs which will be used in next-generation gaming consoles will be able to process geometry like Fermi does as well. And this is the right direction to move in, although it may be a bit premature. Well, someone has to be the first. In case of tessellation it's NVIDIA. Which may seem strange, remembering AMD's generations of tessellators.

NVIDIA released two demos to demonstrate the capabilities of new GPUs: Endless City and Alien vs. Triangles. These two differ by the progressive use of tessellation with a high level of decomposition. For example, up to 600M triangles per second are processed in Endless City.

This demo generates one of the most complex geometry scenes rendered in real-time. Note that buildings are procedurally generated on the GPU from a set of previously constructed objects.

Tessellation provides previously unseen detail levels on objects close to the camera, while remote objects are decomposed into fewer primitives (adaptive tessellation). The demo uses 3D displacement maps, not regular height maps. Also, about 500000 light sources are used, as well as the screen-space ambient occlusion algorithm.

The second tessellation demo is Alien vs. Triangles. It utilizes tessellation to add geometry details to an alien figure, for which several different displacements maps are used simultaneously. The key feature is the use of tessellation to imitate realistic damage.

As stated above, the demo utilizes three different displacement maps: Normal, Spike and Fungus. Another four displacement maps are used as damage maps. When the alien is hit, they are modified procedurally in real time to imitate damage.

But, of course, demos are not the only applications where tessellation is used. Aside from the popular titles like DiRT 2 and Metro 2033, there is the recently released demo of HAWX 2 that also uses tessellation aggressively to render the ground.

As you can see, this is adaptive landscape tessellation. At that the average triangle size, according to NVIDIA, doesn't exceed 18 pixels. This average value will not seriously reduce the efficiency of modern DX11 GPUs. It will be interesting to compare various solutions in this benchmark, when (if) we add it to our list of tests.

Miscellaneous changes

GeForce GTX 580 improvements are not limited to 3D features. Although the difference in power consumption specified by NVIDIA is just 6W (250W of GTX 480 vs. 244W of GTX 580), real-life tests indicate the difference of 20-30W. In other words, the power consumption is about 10% less at about 20% higher performance (we'll see if the latter is true).

The reduced power consumption and improved cooler should reduce noise. The new cooling system features a copper vaporizing chamber and a large, dual-slot turbine fan.

The fan is claimed to be less vibrating and noisy, which is also important. In turn, new rotation speed control algorithms "ease" spin-up under high load. According to NVIDIA, the new cooler is even quieter than that of GTX 285, not to mention that of GTX 480.

Certain changes were introduced into power and temperature monitoring as well. Previously, graphics cards were protected from damage by prohibiting to work at temperatures higher than critical. While now the power consumption of the entire system is monitored.

New hardware monitoring elements monitor amperage and voltage on 12V rails (PCIe, 8/6-pin supplementary connectors). If the power level exceeds the highest, the driver can decrease GPU clock rate in demanding stability tests like Furmark and OCCT.

This limitation only applies to certain applications preset in drivers, games excluded. Today it's only activated in Furmark, on exceeding the capabilities of power rails. In this case, GPU clock rates will be halved.

Conclusions on theoretical information

Obviously, GF110 is an improved and overhauled GF100. A fixed version, so to say. The new high-end GPU is better than the predecessor in several aspects: it has more active execution units, higher clock rates, improved TMUs and z-cull algorithm. But primarily it consumes less power and produces less heat, thus fixing the key problems of GTX 480.

GF110 features all the advantages of the Fermi architecture. This is especially noticeable in the changes made to geometry processing in the graphics pipeline. It is very useful for modern DirectX 11 applications that widely use geometry and tessellation. This is also becoming noticeable in games with support for these respective features, the number of which is growing.

Compared with GF100, GF110 offers the doubled speed of FP16 texture fetching with bilinear filtering, as well as the improved z-cull algorithm. All of this may add a few percent to the rendering performance boost. Moreover, the new cooler and monitoring features add to the consumer qualities of NVIDIA's novelty.

Of course, we shouldn't forget about such important technologies as CUDA, PhysX and 3D Vision. GTX 580 naturally supports all of these.

In general, GeForce GTX 580 is an excellent replacement for GTX 480. Of course, it would be unreasonable to expect great performance improvements, given that the process technology has remained the same. But another +20% to the performance of the previous best single-GPU solution seems quite nice. Especially, considering the aforementioned lower power consumption, heat emission and noise. Now let's see what the performance boost really is in the performance tests.

Write a comment below. No registration needed!