Predictive Power Consumption Model for Compute Intensive Applications in Clustered ARM A53 Embedded Systems

High power consumption has been a concern in x86 architectures. In this same line, alternatives to x86 have been explored in order to have similar or higher ratio of computing capabilities with less power consumption. In order to find a power and cost efficient alternative for supercomputer architectures this paper explores the implementation of a low power ARM cluster based on embedded systems and analyses the cluster power consumption while running MiniMD, a compute intensive molecular dynamics workload. Based on MiniMD data, it is presented a predictive power consumption model for compute intensive applications with a 5% correlation error from real power measurements. The model also correlates within 3% error against Linpack measurements. Linpack is the compute intensive benchmark responsible for the "Top 500 supercomputers" ranking. Finally, by using the created model, power consumption projections for hypothetical cluster hardware configurations are presented. The projections exemplify how in the future, ARM based supercomputers will be a good alternative for reaching better power-performance capabilities.