Something else I noticed about that particular research paper is that the "monochromatic-ness" of the shader makes a huge difference. Most pixel shader operations are done in 3 colors, possibly plus an alpha component. The pixel shading pipelines in a GPU like the RSX have 4-component vector math units to handle these operations. The SPUs in the Cell have similar 4-component vector math units. Ordinarily, one SPU would be able to process one pixel at a time when using full color (or, at best, process 4 pixels at a time, but only one color at a time, requiring 3 or 4 passes for the same net effect). In the research paper, 5 SPUs were able to process 20 pixels at a time (vs. 24 on the 7800 GTX) because their shader was only in one color. For most pixel shader operations, this would not be the case.
In the Killzone presentation, it seems like they've come up with a way to avoid multiple rendering passes, but it still seems to me that the GPU would handle these computations better. But with the vaugeness of the presentation, it's kind of difficult to discuss specifics.
The Insomniac Games paper isn't talking about pixel shaders at all. "SPU Shaders" are an approach to how the game code is designed with respect to using the SPUs to handle tasks. It has nothing to do with this discussion.
