Supercomputing gets personal!
Francis Wray, Concertant
The significant government funding that has supported supercomputing for the last 30 years or more has been based on the premise that it stimulates the economy. This is clearly demonstrated by the impact that intensive processing has had on exploration in the oil and gas industry and simulation in the automobile and aerospace industries. A further economic benefit has been the diffusion of technology into the mass market. Advances in memory, storage, networking and software tools, developed initially for high-end computing, have then percolated down to benefit the PC marketplace. For the last decade, however, the traffic has mostly been the other way with advances in commodity microprocessors enabling the construction of ever more powerful multiprocessor supercomputers. It is clear from discussions at SC08 that the direction of this traffic is about to change again.
GPGPUs, developed for the games market, have developed rapidly, stimulated by significant consumer demand. These devices have also found a niche in supercomputing, where their widespread use has led to significant developments in the way they are programmed. These devices can now be programmed not only as graphics engines, but also as powerful number crunchers. Similarly, other devices such as FPGAs and high core-count devices from Tilera and ClearSpeed are being used for numerically intensive processing. All these developments are now set to filter back to the PC marketplace.
It seems likely that in two years' time, a PC may comprise a powerful multicore processor with an attached coprocessor that will probably be GPGPU-like, but which could also be an FPGA or a high core-count device. Based on manufacturers' announcements, we would expect such a PC to have a performance around two orders of magnitude greater than current machines and would be similarly priced. Such attached processors would not be directly programmable by the ordinary PC user, but they would certainly enable the implementation of extreme applications by professional software developers and even hobbyists, perhaps via some kind of API such as CUDA or OpenCL or others. Application areas, which immediately spring to mind, are image and video compression and decompression and photorealistic rendering for design. These will be complemented by an explosion in existing business applications many of which will have been previously restricted to large clusters, but which would run on this new generation of super PCs. Of course, games will run on these systems, but a whole new set of new popular applications will be enabled allowing much greater productivity.
Supercomputing is about to get much more personal!