The Bigger Picture
It really is not even slightly important how powerful certain GPUs are compared to others. When it comes down to it, FLOP performance is a pretty poor means with which to compare the gaming performance of a graphics card, so any mention of this number in an actual graphics card review should be taken with a grain of salt. What is very important about these numbers, however, is that they are A LOT higher than the numbers produced by modern CPUs. The current best consumer-level desktop processor, the Core 2 Extreme QX6850 from Intel, is capable of somewhere in the neighborhood of 30 GFLOPS. The R600 GPU is capable of 600GFLOPS. This means that the R600 is 20 times more powerful than the QX6850 at Floating Point calculations. The Cell B/E used in the PS3 pushes just over 200 GFLOPS, just a third of the R600’s potential.

Of course by taking a step back from the situation, it is pretty clear that the publicized FP capabilities of these processors are somewhat of a far cry from the realized values. Evidence of this can be found by taking a look at the statistics for the distributed computing program Folding@Home (F@H). As a Quasi Supercomputer, the F@H network ranks pretty high on the list of top overall computing power towards a single cause. This overall computing power is broken down into different F@H clients, which vary depending on the operating system they are used for. Windows-based PCs are currently folding at a rate of 165 TFLOPS. There are 173,075 CPUs currently working to produce this statistic. That means that each CPU is contributing about .95 GFLOPS to the F@H network. PS3 systems are currently folding at a rate of 437 TFLOPS, with 24,162 active systems. Each PS3 is contributing about 18 GFLOPS to the F@H network. The only GPU series that supports F@H is the R5XX from ATI. This core, which is used in Radeon X1000 class graphics cards, features 48 programmable pixel shaders (R580) that are particularly well-suited for molecular dynamics calculations. There are currently 663 ATI GPUs folding at a rate of 39 TFLOPS. This means that each GPU is contributing an average of 59 GFLOPS; a 60x increase over traditional CPUs. So why don’t we see GPUs powering all of our applications if they’re so much better? For an answer to that, we have to go back to very origins of the GPU itself.

1981
As computer graphics began to develop beyond that flashing white underscore that practically doesn’t exist today, the need for a dedicated processor to handle the graphical operations associated with various computing tasks began to emerge. The earliest graphics cards were single color, used 4KB of video memory, and was very effective at displaying text on a screen. All GPUs used on graphics cards produced for a period of about 13 years after 1981 were rather primitive, parallel processors that were used to draw geometry on the screen. In 1995, companies like ATI, S3, and Matrox introduced the first 2D/3D graphics chips that were extremely powerful and far more capable of producing 3D images on the screen than were era CPUs. The whole purpose of the GPU was to take a work load from the CPU and process it so the CPU did not have to. GPUs were designed to be good at floating point operations, such as those common in 3D functions. As such, they really were not very good at other kinds of calculations. That was then.

Twelve years of development later, the GPU has transformed into something that barely resembles the single purpose chip that it was in 1995. The addition of pixel shaders – essentially primitive processors in their own right – and the slow transition (that is now pretty much complete) to a unified shader architecture – such that each shader does not have a specific task and can do whatever needs to be done – has transformed the GPU into, basically, a multi-processor system on a single chip; a supercomputer on a chip if you will. So, although GPUs have been around for more than 20 years, and the idea that GPUs could be used for more than just graphics processing has been around since really 2000, the GPU architecture itself has only been truly hospitable to these kinds of calculations for about four years. So now that it’s here, what are we planning to do with it?