First, most Ars readers are aware that, as a general rule, the TFLOPS metric is a pretty poor way to judge performance. Different instruction set architectures (ISAs) will render the same calculation using different numbers of floating-point operations, sometimes skewing the results significantly when you compare TFLOPS ratings across ISAs. So the TFLOPS ratings given by F@H are fun, but the project itself acknowledges that there are major problems with this metric. In fact, the F@H group's initial numbers for the PS3's TFLOPS rating were wrong, and were revised downward by 50% last night.
Here's a simple analogy to illustrate the logic behind the rankings. The GRE exam has three sections: verbal reasoning, quantitative reasoning, and analytical writing. Let's say that three different students took the exam, and they were then ranked relative to each other based on the average of their three scores.
Because student G is a mathematical genius and an autistic savant who can't do anything but math, he was asked to take only the quantitative exam—an exam that he's insanely good at. So his GRE average reflects his performance on that single exam. Student P excels at both math and vocabulary, so took he took only the quantitative and the verbal reasoning exams. Thus P's GRE average reflects the average of only these two exams. Student C is good at vocabulary, average at math, and a solid writer. He took all three exams, and all three scores contributed to his GRE average.
If we were to rank the average scores of all three students, student G would outclass the other two by a wide margin, student P would come in second place, and student C would be stuck at a far, far distant third place. This is because G (the GPU in the analogy) took only the test that he was insanely good at, P (the PS3) took the two tests at which he excelled, and C (the general-purpose CPU) had to take all three tests.
Ultimately, the TFLOPS/CPU rankings given above align pretty much exactly with the degree of specialization of each type of processor. The GPU is far and away the most specialized of the three, so it sits comfortably atop the rankings. The PS3 has a lower degree of specialization than the GPU, but a significantly higher degree than the general-purpose CPU. Indeed, you could almost use each processor's TFLOPS/CPU score as a sort of "degree of hardware specialization" rating.
The final thing that's worth noting is that the pool of CPUs that make up the "Windows" portion of the client list varies widely, from older Pentium 4 models to brand new Core 2 Duos and everything in between. The GPUs are much more uniform in terms of hardware types (all ATI), and the PS3 is the most uniform of them all. So the PS3-to-PC comparison isn't just apples-to-oranges. It's more like apples-to-citrus.