AGP - The Practice

The Performance Issue

The expected increase in system performance is certainly the most interesting thing to all of us. This question can be answered fairly simple in the first place - there doesn't seem to be any!!! Shocking!! Well, it isn't quite like that, but you have to try hard finding any improvement in performance from using AGP. Using any real world applications or benchmarks won't get you anywhere. It is difficult to believe, but any popular benchmark program will show you the same results with the PCI card as well as with the corresponding AGP card. As to my knowledge there is currently only one benchmark available that can show you the advantage of AGP and its DIME feature, which is the 3D Winbench 97 from Ziff-Davis. Now hold your horses before you start benchmarking with 3D Winbench, because if you are using the benchmark just as it comes, it won't show you any difference between PCI and AGP either. What you will need in addition to 3D Winbench is a special scene called "Station 2 Large Textures Scene". This scene can be downloaded from ZDLabs as 3d97agp.exe. This "Large Scene" got its name because it's using very large textures of up to 6.1 MB. Even new cards with 8 MB onboard memory won't be able to hold the complete textures in their local (onboard) RAM in case you are using high resolutions that need a lot of local memory for frame buffering and Z-buffering.This means that the texture has to be swapped to local memory. Here is where DIME comes in for AGP cards that are able to use it. You can see a benefit of about 200% by using the AGP card, when running the "large scene" at a resolution of 1024x768 . The benefit from AGP decreases whilst using lower resolutions and depending on the amount of local card memory the results will finally be the same at some low resolution, e.g. 640x480 in case of a 8 MB graphic card.

You invoke the "large scene" by selecting 'user scene' in the 'test settings' of 3D Winbench 97. Select 'Stations2LargeTextures.scene' as user scene name and tick the options you want to run the test with. Choose the resolution you want to run the test at and run the test via 'selected' in the 'Run' menu, choosing '3D Scene/User Defined'.

To summarize the findings regarding system performance with AGP graphic cards compared to the corresponding PCI graphics card there are the following things to be considered:

  • I can say that so far there is no real world application available that benefits from AGP's DIME feature.
  • Normal applications, like e.g. business applications don't benefit from AGP at all. In Winstone 97 the Business as well as the HighEnd benchmark are showing identical results between AGP and corresponding PCI card.
  • Current 3D applications are also not benefiting from AGP. The textures used by current 3D games and 3D benchmarks are either not large enough or the games and benchmarks don't run at high enough resolutions for forcing the graphic card to swap parts of the textures outside their local (onboard) RAM into main memory.
  • As new graphics cards are coming with more local (onboard) RAM, the textures can get larger than used currently and still fit into the local RAM. Only applications that use particularly large textures or graphic cards with only 4 MB local RAM or less will see an advantage from AGP, at least for now.