Run-Time Power Modelling in Embedded GPUs with Dynamic Voltage and Frequency Scaling

TL;DR

This paper develops a CPU-based power modeling approach for embedded GPUs that uses performance counters and a unified model to accurately predict power consumption across different voltage and frequency settings, aiding in power estimation when direct measurements are unavailable.

Contribution

It introduces a robust, unified power model for embedded GPUs that accurately predicts power across multiple DVFS points using a single set of coefficients, simplifying power estimation.

Findings

Unified model achieves 5% average error.

Model accurately predicts impact of voltage, frequency, and temperature.

Applicable for power estimation without direct measurements.

Abstract

This paper investigates the application of a robust CPU-based power modelling methodology that performs an automatic search of explanatory events derived from performance counters to embedded GPUs. A 64-bit Tegra TX1 SoC is configured with DVFS enabled and multiple CUDA benchmarks are used to train and test models optimized for each frequency and voltage point. These optimized models are then compared with a simpler unified model that uses a single set of model coefficients for all frequency and voltage points of interest. To obtain this unified model, a number of experiments are conducted to extract information on idle, clock and static power to derive power usage from a single reference equation. The results show that the unified model offers competitive accuracy with an average 5\% error with four explanatory variables on the test data set and it is capable to correctly predict the impact of voltage, frequency and temperature on power consumption. This model could be used to replace direct power measurements when these are not available due to hardware limitations or worst-case analysis in emulation platforms.