Quantifying the energy efficiency of coordinated micro-architectural adaptation for multimedia workloads

Adaptive micro-architectures aim to achieve greater energy efficiency by dynamically allocating computing resources to match the workload performance. The decisions of when to adapt (temporal dimension) and what to adapt (spatial dimension) are taken by a control algorithm based on an analysis of the power/performance tradeoffs in both dimensions. We perform a rigorous analysis to quantify the energy efficiency limits of fine-grained temporal and coordinated spatial adaptation of multiple architectural resources by casting the control algorithm as a constrained optimization problem. Our study indicates that coordinated adaptation can potentially improve energy efficiency by up to 60% as compared to static architectures and by up to 33% over algorithms that adapt resources in isolation. We also analyze synergistic application of coarse and fine grained adaptation and find modest improvements of up to 18% over optimized dynamic voltage/frequency scaling. Finally, we analyze several previous control algorithms to understand the underlying reasons for their inefficiency.