Ancillary Performance
We test a few other non-game areas in our theoretical analysis. Here's the executive summary before we move on to the game performance.PCI Express Bandwidth
On nForce4 SLI, each GPU sits just short of the magical 1GiB/sec mark. Memory bandwidth can affect this test, but it doesn't in this instance.
Memory Bandwidth Performance
Yours truly is also in the middle of engineering a tool that measures GPU memory bandwidth (largely to investigate a HyperMemory board!), and this is the first run out of that tool. Be aware that it's still in a beta state and isn't 100% reliable (colour compression can inflate the result), but it works for this article on the NVIDIA GPUs at least.
Each chip can address its memory at very close to the maximum rate, which for the 7900 GTX is over 50GiB/sec.
Depth Rejection Performance
NVIDIA's GPUs are able to reject pixels before fragment shading, the rasteriser calculating depth for generated fragments, refusing to pass the fragment on to the shader hardware. To test, we have a tool render a scene from front-to-back (depth increases, fragment rejection is easy) and compare its performance to rendering the scene from back-to-front (hard to reject anything at all). If there's a large difference, early Z reject is working.
I think it works, don't you?
