Beyond Dataflow

Abstract

This paper presents some recent advanced dataflow architectures. While the dataflow concept offers the potential of high performance, the performance of an actual dataflow implementation can be restricted by a limited number of functional units, limited memory bandwidth, and the need to associatively match pending operations with available functional units. Since the early 1970s, there have been significant developments in both fundamental research and practical realizations of dataflow models of computation. In particular, there has been active research and development in multithreaded architectures that evolved from the dataflow model. Also some other techniques for combining control-flow and dataflow emerged, such as coarse-grain dataflow, dataflow with complex machine operations, RISC dataflow, and micro dataflow. These developments have also had certain impact on the conception of highperformance superscalar processors in the “post-RISC” era.