The modern computer seems to be used more now as a vehicle for connecting to the rest of the world than the instrument of computational aid that it was originally intended to be. Although it is probably not used for such a purpose as often as it should be, the computer sitting in front of you right now is quite capable of carrying out complex mathematic and scientific calculations at speeds that, just 10 years ago, would have seemed impossible. There are, however, computers out there that are used strictly for the computational means for which they were developed. These machines, often housed in expansive, air conditioned rooms in the fine establishments of the academic world, differ in one rather fundamental way from the personal computers that reside in our homes.

The parallel architecture used in modern supercomputers is actually not much different than the technology used in most desktop level PCs. Indeed, modern supercomputers are really just clusters of SIMD (Single Instruction, Multiple Data) processors, not unlike those used in the consumer market. The aforementioned “cluster” however, is what allows supercomputers to achieve levels of massive parallelism not found in the personal computer. Interconnects and shared memory architectures implemented in the top modern supercomputers (currently IBM’s Blue Gene/L, housed at the Lawrence Livermore National Laboratory) are more reminiscent of sever farms and network rendering setups than any modern personal computer. The key here is not only the number of processors involved, but the way in which they collaborate.

The fastest x86 (instruction set used by essentially all personal computers) processors from AMD or Intel, multi-processor server configurations, are capable of just about 60 GFLOPS. The measure of ‘FLOP’, or Floating Point Operations per Second, is the most commonly used means of determining the shear computational power of computing systems. The SI system of prefixes applies to FLOPS, and as such a GFLOP is simply a billion FLOPS. By contrast, the fastest modern supercomputers are capable of 280 TFLOPS, or roughly 4,500 times more FLOPS than the fastest PC. When Sony’s Playstation 3 game console was announced, it was well-publicized that the system was capable of pushing 2 TFLOPS. The Cell Broadband Engine processor used in the PS3 itself was capable of a little over 200 GFLOPS, still about four times faster than the most capable PC.

When you look at the math for the supposed performance capability of the PS3, you notice that somewhere along the line 1.8 TFLOPS of information is unaccounted for. According to Sony marketing material, this missing horsepower manifests itself in the Reality Synthesizer, also known the RSX graphics processor supplied by NVIDIA. A graphics processor pushing 1.8 TFLOPS? You know what they say: if it sounds too good to be true it probably is. Independent tests, and even tests conducted internally from Sony, have revealed that the actual floating point performance of the RSX is more in the 300 GFLOPS neighborhood. This was certainly not the kind of news that champions of the PS3 had wanted to hear, but the number was still an extremely significant one. The RSX in the PS3, for all intents and purposes, is just a slightly modified GPU not unlike those found in most gaming PCs. In fact, the RSX is only half as powerful many of the very fastest GPUs used in today’s gaming PCs. The R600 GPU used in AMD’s Radeon HD2900XT graphics card is capable of just over 600 GFLOPS, with NVIDIA’s G80 not far behind. Two of these cards in CrossFire or SLI breach the TFLOP range.