Program-driven fine-grained power management for the reconfigurable mesh

The reconfigurable mesh model for massively parallel computing has recently been rediscovered and proposed as the basis of a practical many-core architecture. With this paper, we are the first to study the power and energy requirements of reconfigurable meshes. We present two methods to reduce the power consumption, exploiting characteristics of typical reconfigurable mesh algorithms. We extend our previous reconfigurable mesh architecture and the corresponding programming language ARMLang to support program driven power management. On two sample applications, we discuss model-based power estimations and compare the estimates with power measurements on FPGAs. Our program driven power management methods are shown to be effective. For one application we report a reduction in power and energy consumption by 21.09%.