Nvidia GeForce GTX 580
As we’ve already said above, the GeForce GTX 580 is meant to become Nvidia's new top-of-the-line solution, the flagship of the entire Fermi line-up, the proof of the company's technological superiority in graphics card development, and the winning trump against the upcoming Radeon HD 6900. Let's first see how the specifications of the new card have improved over those of the GeForce GTX 480.
The first thing we can see is that the clock rates of both the graphics core and memory have grown up. Other things being equal, this fact alone would make the new card considerably faster. The die has become smaller by 9 square millimeters. This is an expected outcome of the optimizations the engineers have made to the successor of the GF100. What is somewhat unexpected is that the new GPU is said to incorporate 200 million transistors less than its predecessor. This is not a trifle after all and 200 million transistors should have made up quite a big portion of the core. Well, we don’t think that’s some kind of a sensation. Although all the official sources say that the GF100 contains 3.2 billion transistors and the GF100, only “about 3 billion”, and although the popular diagnostic tools like GPU-Z report the same values, no one, except for the GPU developers themselves, can know the exact number of transistors. The software tools simply report what is written into their databases. It is in fact impossible to learn the exact number of transistors in a given chip. The GF110 may be indeed slimmer than its predecessor due to optimizations, but the difference may be something like "3.09 billion now against 3.15 billion before" and that would sound far more realistic than the loss of 200 million transistors at a time, the two GPUs having the same architecture and the same number of functional subunits.
Next we see 512 active ALUs. This is the same number of ALUs as were physically present in the GF100 but some of them were not activated. The structure of the new graphics core has remained intact:
Each of the four graphics processing clusters incorporates four stream multiprocessors. Each of the multiprocessors consists of 32 general-purpose execution cores and is allotted four texture-mapping modules.
There were rumors that Nvidia would increase the number of texture-mapping processors in the GF110 up to 128 to correct the misbalance of the GF100's TMU subsystem, but the new core still has only 64 TMUs. However, the developers have revised their architecture so that the TMUs could filter all texture formats at full speed whereas the GF100 used to filter FP16 textures at a half speed. This is similar to the texture processors of the G80 and G92 chips: the G80 had two filter units for each fetch unit whereas the G92 had the same number of both types of units. As a result, the G80 had an advantage when doing anisotropic filtering. Things must be more sophisticated with the GF100 and GF110 and we won't speculate on them since we don’t know the details of the TMU design in the Fermi architecture. Let’s just take it for granted that if the GF110 can process textures of all formats at full speed, this makes it superior to its predecessor in modern games which make wide use of the FP16 texture format.
Besides the improved texture-mapping units, Nvidia has also optimized the occlusion culling algorithms. Coupled with the 512 active ALUs, this should guarantee the GeForce GTX 580 the top place among the fastest single-GPU graphics cards and, perhaps, make it strong enough to beat the long-time leader Radeon HD 5970, which is a dual-processor solution.
As for tessellation, there were no problems in this area with the GF100 and even with the GF104. The GF110 is even going to be somewhat faster when doing tessellation because all the 16 geometry processing units called PolyMorph Engines are active in it whereas its predecessor had only 15 (and these engines are also clocked at a higher frequency in the new GPU). AMD may have something to say on this issue with its upcoming Radeon HD 6900, but so far Nvidia’s solutions are superior in this respect. Although there is an opinion that excessive tessellation may be unnecessary or even harmful, it is better to have a graphics card capable of coping with scenes like in H.A.W.X. 2 that have up to 1.5 million polygons per frame than not to have one. Every argument against too much tessellation looks lame compared to a real hardware device capable of doing so much tessellation. On the other hand, Nvidia's claims that its solutions are 20 times as fast at tessellation as their opponents should not be trusted blindly, without a practical check. Besides, the biggest problem with tessellation is that there are still too few games available that make use of it.
Summing everything up, the Nvidia GeForce GTX 580 looks a well-balanced solution of the premium class. Its recommended price equals that of the GeForce GTX 480 at the moment of the latter’s announcement, i.e. $499. This makes the new card more appealing than the Radeon HD 5970 which still cannot be found for less than $600.