The SVIST (Shader Visualisation Technique ) test indicates the type of pixel shader that's being used. When I ran SVIST on the 9800 Pro, I would say that 50-60% of the output was based on PS 2.0. I'm not sure if the test provides ending statistics since I cancelled it after 3 minutes or so.
This information is from the AM3 documentation on shaders.
The effects and their variations are currently implemented by 190 different vertex shaders and 39 pixel shaders. These numbers of shaders in combinations with the various non-pixel shader rendering techniques for DirectX 7 class hardware allow for testing the rendering hardware and their resource management to its limits. In the following table you find an AquaMark3 shader overview.
Number of shaders:
VS1.1 (DX8) : 190
VS2.0 (DX9) : 3
PS1.1 (DX8) : 30
PS1.4 (DX8) : 5
PS2.0 (DX9) : 4
Average number of insturctions per shader:
VS1.1 (DX8) : up to 70
VS2.0 (DX9) : up to 50
PS1.1 (DX8) : up to 13
PS1.4 (DX8) : up to 13
PS2.0 (DX9) : up to 40
4.1.2 Floating Point Numbers In Shaders
AquaMark3 applies the whole range from FP64 to FP128 floating point accuracy. We think the developer is still in charge to decide which accuracy should be considered for a certain effect to optimize the rendering efficiency.
In general, every effect with floating point numbers covering a wide numerical range and the need to be represented by an exponent (as world space coordinates per pixel) benefits from a high internal precision (FP128). That means mainly special effects which use geometric information inside the pixel shader (as high dynamic range illuminated normal mapping with a self shadowing term) need high precision. AquaMark3 does not require functionality in this respect.
Nevertheless we use high internal precision for the texture coordinate iterators because AquaMark3 utilizes effects which calculate texture coordinates on the basis of world space positions (caustics, fog, detail maps etc.). Besides, the majority of AquaMark3s effects are implemented in less accuracy and make availible precision hints to the API.
There's a thread in the AM3 forum where this was being dicussed. It appears that AM3 will not "fall back" to an earlier shader version in the TRISCORE (default benchmark) run.