We are on the cusp of a new era in computing. As microprocessors become more powerful and have more cores per die, there is less need for additional general purpose computational power. On the desktop, the computational load is primarily graphics, image processing or encoding/decoding of music and video. These tasks are computation heavy and branch/logic light, much like traditional supercomputing. As a result, the major microprocessor producers have been moving toward more floating point computational power in their processors. IBM produced the Cell processor, a powerPC core with 8 simpler vector processing cores, which is the workhorse for the first petaflop computer. Obviously, it is the fastest in the world. The top two graphics processing companies, nVidia and AMD, are also becoming more concerned about developing programming tools to allow the computation power of their graphics processors to be used for purposes other than graphics. Finally, Intel will be extending the x86 code base for vector processing when they produce an x86 based graphics accelerator codenamed Larrabee.
Computation power has always been important in research. Simulating nuclear devices is computation intensive, so the DOE has always had a top notch system in New Mexico. However, a new field is opening up that requires much more, biology. Specifically, the task of understanding protein folding and interaction. Stanford’s Folding @Home program asks people to borrow the processing power of their computers that their not using to do protein folding calculations. From the beginning, the PS3, which is powered by IBM’s Cell processor has been a strong contributor to the program. Recently, they have also developed a client in Cuda, the nVidia proprietary language which promises to bring the substantially higher processing power of GPUs to help solve the protein folding problem.
The only problem with Folding @Home is that the processing power of individuals is so small that it is really not possible to simulate a significantly long folding sequence. At least that is the claim made by D. E. Shaw. There is also an article in the New York Times, which is less technical.
More or less, Shaw’s argument is that a dedicated supercomputer is needed and he can produce a specialized ASIC that will do the job 1000 times faster than the processors used in current supercomputers in about 5 years. Unfortunately, while there will be an approximately 10x shrink in that time, supercomputers will be in excess of 100 times more powerful. Possibly 1000. This is because all but one of the top supercomputers is powered by either Intel Xeon processors, AMD Opteron processors or IBM Power processors. The emergence of the new Cell based system IBM built for the DOE and deals by Intel with Cray and Dreamworks suggest that mainstream supercomputing will no longer be driven by just general purpose CPUs, which aren’t very efficient at raw computing. Larrabee is a big part in this, as will Cell and Cuda. D. E. Shaw’s Anton is going to be yet another specialty chip that will be marginalized by higher volume processors.