For the last few years, High Performance Computing (HPC) systems have become highly heterogeneous through an unyielding integration of multi-core processors (CPUs) and accelerators. Among accelerators, Graphics Processing Unit (GPUs) have emerged as ideal co-processors for high-throughput applications, due to their enormous computational power and energy efficiency. However, these GPU-based systems still consume a lot of energy. In order to provide Exaflops, these extreme-scale systems must be optimized for "performance-per-watt". Similar to any optimization process, it is first required to understand how power is dissipated into general purpose processors. In this paper, we propose a macroscopic approach for analyzing the GPU power consumption through the benchmarking of a memory bound application kernel 1. This approach offers great insights into inherent characteristics of GPU power behavior under different configurations.