How GPU DVFS Boosts Efficiency: Concepts, Modeling, and Future Directions

Background: Reducing power consumption in compute systems can be achieved by low‑power techniques that keep performance unchanged or by reducing processor count; DVFS belongs to the former, adjusting frequency and voltage at runtime to match workload demand.

Basic Concepts

DVFS

DVFS consists of Dynamic Voltage Scaling (DVS) and Dynamic Frequency Scaling (DFS). Two principles underlie it: (1) dynamic power is proportional to the square of voltage, so lowering voltage sharply cuts power; (2) performance is roughly linear with clock frequency, so reducing frequency lowers performance, but voltage reduction can mitigate the loss.

GPU Boost 4

NVIDIA GPU Boost 4.0, the fourth generation of GPU clock technology, converts available power headroom into higher performance and succeeds GPU Boost 3.0. It introduces three new features: no strict temperature limit, user‑editable temperature thresholds, and an NVIDIA OC scanner.

Power‑Optimization Model

Reference: “NVIDIA‑General GPU Power Estimation Model”. The authors construct a GPGPU power‑estimation model for the Kepler architecture using hardware performance‑counter events.

Model Construction Steps

Select power‑relevant performance events and collect GPU power values together with the corresponding counters.

Design a GPU power‑estimation model based on a Back‑Propagation Artificial Neural Network (BP‑ANN).

Implement and evaluate the GPGPU power model.

Performance‑Counter Selection

From 141 hardware counters, 12 (mostly memory‑related) were chosen because they correlate closely with power consumption.

Model Accuracy

For eight benchmark applications the model error stays below 5 %, showing that a hardware‑counter‑based GPGPU power model can accurately predict runtime power.

Future Directions

Intelligent DVFS: apply machine‑learning techniques to improve DVFS decision accuracy.

Multi‑GPU optimization: coordinate DVFS across multiple GPUs for global power savings.

Technology fusion: combine DVFS with task scheduling and load‑balancing to further raise system energy efficiency.