NVIDIA Quadro 4 980 XGL (AGP 8x) | NVIDIA Quadro 4 900 XGL |
 |  |
by FastSite
12/14/2002 | 12:00 AM
Not so long ago NVIDIA announced their first professional graphics card aka Quadro 4 980 XGL. As you can easily guess, it differs from the predecessor, Quadro 4 900 XGL, only by the implemented support of the new AGP 8x bus, while the working frequencies of the product remained the same: 300MHz for the chip and 650MHz for the memory. If you take a closer look at the pictures available on NVIDIA's official web-site, you will definitely get the impression that the cards follow the same reference design, however, it is not true. Have a look here:
| NVIDIA Quadro 4 980 XGL (AGP 8x) | NVIDIA Quadro 4 900 XGL |
| |
With a naked eye you can notice the differences.
Quadro 4 900 XGL is based on the GeForce4 Ti4600 layout, while Quadro 4 980 XGL has been significantly redesigned. However, I would like to repeat once again that the chip and memory working frequencies remained the same. As well as the memory chips access time, which remained equal to 2.8ns.
Before we pass over to the actual tests, I would like to draw your attention to the info placed on NVIDIA's web-site, which should give us some idea of the graphics cards performance:
In other words, NVIDIA claims that the performance in 3dsmax should get at least 5% higher due to the shift to AGP 8x interface. Well, let's ask ourselves first: what improvements can we expect the shift from AGP 4x to AGP 8x to bring us? The theoretical part has been described in great detail in our recent NVIDIA NV18/NV28 and AGP 8x Investigation. Therefore, we would like to skip this part here, as you may always check that article for more details. Instead, let's pass over to the most important innovation: the increase in the memory bus bandwidth introduced with the launching of the new AGP interface. And how could it help improve the situation?
In 3dsmax viewports the textures are replaced with the miniatures of the size varying from 128x128 to 1024x1024 pixels (the drivers of some professional graphics cards allow setting bigger sizes as well). This implies that all miniatures like that will definitely fit into the graphics card local memory. And the geometric data transfer rate will hardly be noticeable even with the "non-naked" eye.
Well, let's say a few words about our testbed and methodology and then pass directly to the performance.
Our test system was configured as follows: to perform the tests I decided to use a set of benchmarks recommended by the 3ds max 5 developer company. When I ran the tests in viewports, the screen resolution was set to 1280x1024@32bit, the results are all given in frames per second. The final rendering tests were run with the screen resolution set at 800x600 and the results are given in minutes (this means that the smaller is the value, the better is the performance).
Here is our testbed configuration:
We used the following software:
Note: I decided to take a SiS648 based mainboard as a test platform instead of VIA KT400 based one, because the latter proved slower in AGP 3.0 8x mode than in AGP 2.0 4x mode, which has already been described in detail in the NVIDIA NV18/NV28 and AGP 8x Investigation.
This benchmark checks if the graphics card is capable of refreshing more than one viewport in 3ds max simultaneously. There is some animation played in all viewports, and to load the graphics accelerator even more each viewport is displayed in a different mode: from Wireframe to Smooth + Highlights.
Here the graphics cards have to work with animation in a single viewport. During the test the camera is flying above the rocks and hills of the moon surface landscape, which is built by 400 thousand polygons displayed in Smooth + Highlights mode.
This test is none other but the same moon surface picture from the previous benchmark. However, besides the landscape itself, we now have some flying objects, like pace crafts or planes.
This benchmark deal with the processing of multiple light sources. Since most graphics cards do not support more than 8 light sources, this test as well as the next two work with 8 lights of different types. Here we will have 8 SpotLight light sources, which move and light some geometrical object.
We should point out that imitating the effect made by SpotLights is a much more resource-hungry process than the imitation of Omni or Directional lighting.
Here we have the same object, but this time it is lit by 8 Directional lights. Directional lights in 3ds max 5 are the fastest unlike the previous package version.
Again we've got the same object and 8 light sources. But this time these are all Omni lights, which involve average resources between SpotLight and Directional lights described above.
This scene boasts "easy" geometry and a couple of light sources. It is displayed in a single viewport on the entire screen thus being an excellent test of the rasterizing speed in Smooth + Highlights mode.
This test is intended to show how fast the graphics cards are when it comes to multiple textures processing. The file contains a lot of textures and very little geometry.
This benchmark emulates the work on the game level, as it contains both: sufficient geometry and numerous textures. The animation is arranged in such a way that the entire scene could be displayed completely.
This test reveals the ability of the graphics accelerators to display textures on the deforming geometry.
This test is aimed at showing what the graphics cards are capable of in terms of transparent textures processing.
The new 3ds max 5 features not only the transparency remaining from the previous version, which is imitated by dithering:
But also the 'real" transparency implemented by blending the pixel color of the overlapping objects:
You can shift between the transparency modes in the viewport control panel:
Of course, you can guess that the more correctly implemented transparency will be slower.
Here the camera flies through the rocks and hills of the moon surface landscape built of 400 thousand polygons, i.e. the scene is the same as in Benchmark 2, actually. However, the picture is displayed in the Wireframe mode.
This is the Benchmark 3 scene in wireframe mode:
As we see, the results are almost identical. These results were obtained for 128x128 pixel miniatures. After that I thought that AGP 8x might reveal its advantages once the miniature size gets bigger, that is why I increased the miniature size up to the allowed maximum of 2048x2048 and ran Benchmark 9. Here are the results obtained:
This is again a complete similarity. After that I again ran all the texture benchmarks with all the miniature settings up to the maximum, and in all cases without any exceptions I got the same results. Well, there is only one conclusion that I could draw here: AGP 8x doesn't offer us any advantages over the previous interface so far.