Empirical and theoretical lower bounds on energy consumption for networks on chip

This paper focuses on the network on chip of multi-core systems and proposes empirical and theoretical lower bounds on the energy consumption of applications. The empirical method consists of an linear programming model that simultaneously reduces communication distances and network traffic. When applied to standard benchmarks, our method shows that locality exploitation can lead to 50% energy reduction on average compared to no optimization. The theoretical lower bound is based on the Rent's rule model from VLSI design, and is obtained analytically from the communication graph structure of applications. The theoretical results show excellent agreement with the empirical lower bound.