This paper surveys system-level dynamic power management (DPM) techniques. DPM is a design methodology that dynamically reconfigures systems to provide requested services and performance levels with minimal active components or load. It achieves energy-efficient computation by selectively turning off or reducing the performance of idle system components. DPM is used in various forms in portable and some stationary electronic designs, though its full potential is often unexplored due to complexity in interfacing heterogeneous components. The paper first describes how systems employ power-manageable components and the impact of dynamic reconfiguration on power consumption. It then analyzes DPM implementation issues and recent initiatives in standardizing hardware/software interfaces for software-controlled power management. The paper models a power-managed system as a set of interacting power-manageable components (PMC's) controlled by a power manager (PM). PMC's are modeled as black boxes, with focus on how they interact with the environment. PMC's can have multiple operation modes spanning the power-performance tradeoff. Transitions between modes have costs in terms of delay or performance loss. A power state machine (PSM) is used to model PMC's, with states representing different power-performance tradeoffs. The paper discusses the characteristics of PMC's, including their ability to switch between high-performance high-power and low-power low-performance modes. The paper reviews different approaches to DPM, classifying them into predictive schemes and stochastic optimum control. It discusses the importance of workload prediction and the challenges of managing power in systems with varying workloads. The paper also addresses the applicability of DPM, noting that it depends on the system-workload pair. It evaluates the break-even time for inactive states and the potential power savings achievable through DPM. The paper discusses predictive techniques for DPM, including fixed timeout policies and threshold-based predictors. It also addresses the limitations of these techniques, such as performance penalties on wakeup and the need for adaptive techniques to handle nonstationary workloads. The paper introduces stochastic control as a more general approach to policy optimization, addressing the uncertainties inherent in system behavior. The paper concludes by highlighting the importance of DPM in achieving energy-efficient computation in electronic systems.This paper surveys system-level dynamic power management (DPM) techniques. DPM is a design methodology that dynamically reconfigures systems to provide requested services and performance levels with minimal active components or load. It achieves energy-efficient computation by selectively turning off or reducing the performance of idle system components. DPM is used in various forms in portable and some stationary electronic designs, though its full potential is often unexplored due to complexity in interfacing heterogeneous components. The paper first describes how systems employ power-manageable components and the impact of dynamic reconfiguration on power consumption. It then analyzes DPM implementation issues and recent initiatives in standardizing hardware/software interfaces for software-controlled power management. The paper models a power-managed system as a set of interacting power-manageable components (PMC's) controlled by a power manager (PM). PMC's are modeled as black boxes, with focus on how they interact with the environment. PMC's can have multiple operation modes spanning the power-performance tradeoff. Transitions between modes have costs in terms of delay or performance loss. A power state machine (PSM) is used to model PMC's, with states representing different power-performance tradeoffs. The paper discusses the characteristics of PMC's, including their ability to switch between high-performance high-power and low-power low-performance modes. The paper reviews different approaches to DPM, classifying them into predictive schemes and stochastic optimum control. It discusses the importance of workload prediction and the challenges of managing power in systems with varying workloads. The paper also addresses the applicability of DPM, noting that it depends on the system-workload pair. It evaluates the break-even time for inactive states and the potential power savings achievable through DPM. The paper discusses predictive techniques for DPM, including fixed timeout policies and threshold-based predictors. It also addresses the limitations of these techniques, such as performance penalties on wakeup and the need for adaptive techniques to handle nonstationary workloads. The paper introduces stochastic control as a more general approach to policy optimization, addressing the uncertainties inherent in system behavior. The paper concludes by highlighting the importance of DPM in achieving energy-efficient computation in electronic systems.