Specs and synthetics
Before we get to play any games, we should stop and look at the specs of the various cards we're testing. Incidentally, the numbers in the table below are derived from the observed clock speeds of the cards we're testing, not the manufacturer's reference clocks or stated specifications.

Theoretically, the GeForce GTX 285 has an edge in almost every category over any other single-GPU graphics card, with the exception being the Radeon HD 4870's slightly higher peak shader compute rate. In fact, the GTX 285 even beats the Radeon HD 4850 X2 at theoretical fill and texture filtering rates.

Interestingly, the GTX 295 has higher peak fill and filtering rates than the 4870 X2, but the 4870 X2 has slightly higher memory bandwidth and decisively more shader arithmetic potential. Then again, we've found that current GPU architectures don't always behave as their specifications might suggest. Perhaps some synthetic benchmarks will help sort things out for us.

None of the cards reach anything close to their theoretical peak color fill rates in this synthetic test. We've found that memory bandwidth often tends to be the limiting factor here. The results do seem to line up accordingly, roughly speaking, with the GTX 285 ahead of both the Radeon HD 4870 and the 4850 X2, while the GTX 295 trails the 4870 X2.

As far as I can tell, the units for 3DMark's texture fill rate test are just completely horked, and FutureMark has shown no interest in fixing this problem. We'll probably switch to a different benchmark utility as a result eventually, but for now, I believe we can use these numbers for relative comparisons, regardless. What these results show us is that the Radeon HD 4800 series has a substantial advantage over the current GeForces in delivered texture fill and filtering rates. The GeForce GTX 285 scores just a little lower than the Radeon HD 4850, and the GTX 295 is behind the 4850 X2.

In spite of the Radeons' theoretical advantage in arithmetic rates, the GeForces tend to be faster in three of the four 3DMark shader tests. In the GPU particles and GPU cloth tests, a second GPU doesn't seem to help and sometimes even hurts performance. That's the nature of the beast with multi-GPU schemes. Not all games or applications scale well across multiple graphics processors.