Designing an MPSoC architecture with run-time and evolvable task decomposition and scheduling: A neural network case study

Decomposition of programs into concurrent tasks and scheduling them among computational recourses are two major problems in hardware and software developments of multiprocessor systems. This paper presents a novel homogeneous multiprocessor architecture, in which a hardware core performs these two jobs at run-time using genetic algorithm. This core looks for an efficient decomposition and scheduling scheme for the running application based on available computational resources. The main novel feature of this system is its capability of executing uni-processor sequential programs directly by cooperation of all available processors. This system is called EvoMP (Evolvable MultiProcessor) and recently introduced in detail by the authors (in another literature). This paper presents a brief description of the operational and architectural aspects of EvoMP and studies applicability of this platform for neural network applications.