ABSTRACT
Architects of general-purpose processors face a herculean task: to design a processor that will run every application fast. However, applications vary widely in instruction mix, frequency, and patterns of data access, input and output bandwidth requirements, etc. A designer may incorporate elaborate and space-consuming circuitry that simulation shows will dramatically improve performance for one application but has no effect on another-or worse, slows it down. For example, designers will normally incorporate as large a cache as space allows on a die, since cache speeds up most applications; however, the data cache adds nothing to the performance of an application that copies data from one place to another,
Programmable hardware can be used to build systems in which the circuitry matches the structure of the problem. In particular, inherent parallelism in problems, which a general-purpose processordespite multiple "go-fast" enhancements-cannot exploit, can be exploited in a system in which multiple circuits are used to speed up the computation.
