Vector power versus parallel power
Supercomputers can do in a matter of hours what hundreds of ordinary computers would take months to do. In fact, supercomputers perform tasks considered virtually impossible for ordinary computers.
The first supercomputers had large memory capacities and performed complex computations using vectorisation software. In vector processing, mathematical equations are solved as whole chunks or arrays, instead of being broken up into simple units.
As single-processor vector supercomputers were pushed to greater speeds, they reached what scientists call "speed-of-light limitations" -- a point where they cannot work faster. It became difficult to reach the vectorisation levels required to extract maximum performance. Another problem was that vector computers generated heat, making expensive cooling systems necessary.
These factors led to a different way of computing -- parallel processing, where a cluster of processors operating simultaneously shared the work. The basic unit of parallel supercomputers is the microprocessor that works with other microprocessors at a single problem. There are two types of parallel computers -- the coarse-grained type, with two to eight processors, and the fine-grained type, with upto hundreds of processors. There are also "massively parallel" computers.
Along with low cost and the ability to operate with simple air-cooling, parallel supercomputers offer "dynamic configurations." For example, an 8-node supercomputer (one with 8 parallel processors) can be upgraded to 16, 32, 64, 128 or 256 nodes, depending on need. Theoretically, the computing power goes up by the same factor as the number of processors. The most developed parallel systems in the world can accommodate thousands of processors. Vector computers sometimes use parallel processors, though not in the same way as parallel supercomputers.
Assessing supercomputer performance is not easy. Because they spend most of the time in mathematical calculations, an important parameter of performance is the number of floating point operations per second (flops) they can do. A floating point calculation is one that involves non-integers. This value is measured in megaflops (one million flops) and gigaflops (one billion flops). Manufacturers are now aiming at teraflops (one trillion flops).
For gross comparisons, peak speed is considered. However, for practical considerations, sustained speed is taken into account. But sustained speed, which can be as low as one-third of peak speed or even less, varies greatly for different programmes. The yardstick often used to assess a supercomputer is the time it takes on an application.