Hardware for high-performance computing: abstract progress, painful consolidation

The history of high performance computing can be displayed in several distinct shapes, each defined by the choice of light used for illumination. Two steady, smooth curves appear if high performance computing is viewed in the abstracted light of technological progress. Both curves trace changes over time roughly, the two decades since Seymour Cray introduced the technology (as well as the name and concept) of vector intensive supercomputing. One curve, representing hypothetical gross performance measured in floating point operations per second (flops), rises majestically toward the brink of teraflops performance. The other, measuring net purchase price per unit of performance, dives gracefully in the opposite direction. The two, considered together, delineate the present happy state in which unprecedentedly high performance is available at prices that were only a dream a few years ago.