Quantifying the insanity
Shader arithmetic is becoming ever more important as games take better advantage of GPU programmability, but basic pixel fill rates and texturing capabilities remain an important component of overall performance. Delivered performance in these categories is also very much tied to memory bandwidth, so we'll take a look at that, as well. Here are the theoretical peak numbers for single graphics cards; optimally, these numbers would double in SLI or CrossFire, with perfect scaling.

Core
clock
(MHz)		 Pixels/
clock		 Peak
fill rate
(Mpixels/s)		 Textures/
clock		 Peak
fill rate
(Mtexels/s)		 Effective
memory
clock (MHz)		 Memory
bus width
(bits)		 Peak memory
bandwidth
(GB/s)
GeForce 7900 GTX 650 16 10400 24 15600 1600 256 51.2
Radeon X1950 XTX 650 16 10400 16 10400 2000 256 64.0
GeForce 8800 GTS 500 20 10000 24 12000 1600 320 64.0
GeForce 7950 GX2 2 * 500 32 16000 48 24000 1200 2 * 256 76.8
GeForce 8800 GTX 575 24 13800 32 18400 1800 384 86.4

Note that the GeForce 7950 GX2 as listed already includes two G71 GPUs with their associated memory subsystems. You've got to then double those numbers for quad SLI configs. Nevertheless, the GeForce 8800 GTX has more memory bandwidth than the GX2, so dual GTXs in SLI will have more available memory bandwidth than a quad SLI rig. Yikes.

The GeForce 8800 GTS, meanwhile, doesn't compare favorably to the GeForce 7900 GTX in terms of pixel and texel fill rates, but you might suspect that won't be an issue when it comes time to run the latest games. Let's see how well these single and multi-GPU configs deliver on their theoretical promise in a synthetic benchmark.

Both GeForce 8800 SLI systems come close to their theoretical peaks for multitextured fill rate here, and those are very high indeed. They're not the highest of the bunch, though. The quad SLI rig is the fastest, and the GeForce 8800 GTS SLI setup trails the GeForce 7900 GTX in SLI. Trouble is, we're about to go prove that doesn't really matter.