1991 | OriginalPaper | Buchkapitel
Parallel Computers: Toys or Tools?
verfasst von : Andreas Reuter
Erschienen in: Supercomputer ’91
Verlag: Springer Berlin Heidelberg
Enthalten in: Professional Book Archive
Aktivieren Sie unsere intelligente Suche um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
The talk starts with the observation that there often is a huge discrepancy between the claimed peak performance of a system and the actual performance experienced by the average user of that system running his/her program. The two figures can be two orders of magnitude apart for classic vector-oriented multi-processors, and the chances for that happening are even higher for parallel machines of all types. So a user buying a computer that is announced at 10 GFLOPS peak performance may end up having a meek 100 MFLOPS, as observed by his specific applications. This is particularly embarrassing since to get 100 MFLOPS one does not need a supercomputer; two workstations for $ 40K in total will do. Now this holds for the well-understood domain of vector processors; what should anyone expect from a machine with 256K processors, for which the vendor claims, say, 20 GFLOPS?