On Energy Nonproportionality of CPUs and GPUs

Energy proportionality (EP) means designing a system that consumes energy proportional to the amount of work it performs. For an EP system, optimizing an application for performance also optimizes the application for total energy. Energy-proportional multicore CPUs and graphics processing units (GPUs) are fundamental to addressing the grand technological challenge of energy efficiency in Information and Communications Technology. In this work, we formally propose strong and weak notions of EP for modern microprocessors. Multicore CPUs were experimentally found to violate both strong and weak EP. This work presents the first attempt at a theoretical analysis to explain the behaviour. GPUs are carefully designed with on-chip resources primarily dedicated to achieving high arithmetic throughput rather than caching and flow control. Consequently, the mainstream view is that GPUs exhibit strong and weak EP. However, GPUs were experimentally found to violate strong EP. In this work, we experimentally study the weak EP of an Nvidia K40c GPU and an Nvidia P100 PCIe GPU using a specially designed matrix multiplication application. We show that both the GPUs also breach weak EP, which presents an opportunity for bi-objective optimization of the application for dynamic energy and performance. By analyzing the Pareto fronts of dynamic energy and performance for a wide range of workloads, the maximum dynamic energy savings are up to 18% while tolerating a performance degradation of 7% for Nvidia K40c GPU and (50%,11%) respectively, for Nvidia P100 PCIe GPU.