A decoupled access/execute architecture for efficient access of structured data

The authors introduce a decoupled access/execute architecture (DAE) that decouples the data access tasks from the data computation tasks and overlaps execution of the two types of tasks. Such overlapping allows at least part of the data access overhead to be hidden. A method for maximizing the decoupling of structured data references from the data computation tasks and a DAE architecture with adequate mechanisms to support and take advantage of such decoupling are presented. The simulation study showed that the access mechanisms provided by this architecture lead to substantial reductions in the data computation units tall time and to significantly more efficient use of the data bandwidth.