GPU Power Measurement via API Utilization and Frequency Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Measuring GPU power consumption in embedded systems is challenging due to the lack of dedicated measurement equipment and the system overhead caused by repetitive access to device drivers for power sensor registration.
Innovation Solution
A power modeling method that calculates a power coefficient using API-controlled GPU utilization and frequency measurements, allowing for real-time power estimation through synchronized data analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a power sensor is installed in the GPU to measure power consumption directly, then measurement accuracy is improved, but system overhead increases due to repetitive access to device drivers
Solution Approach 1:
The patent performs preliminary actions by measuring and storing GPU utilization and driving frequency data during system operation, then uses this pre-collected data to calculate power consumption through a power model. This eliminates the need for repetitive driver access during actual power measurement, reducing system overhead while maintaining measurement accuracy.
Solution Approach 2:
The patent introduces an intermediary approach by using a power model that relates GPU utilization and driving frequency to power consumption. Instead of directly measuring power through the device driver, the system uses this mathematical model as an intermediary to calculate power consumption from easily obtainable performance metrics, thereby reducing driver access overhead.
2Productivity
If a power sensor is installed in the GPU, then real-time power measurement capability is improved, but device complexity increases
Solution Approach 1:
The patent extracts the power measurement function from the physical hardware domain and relocates it to the software/domain of performance monitoring and mathematical modeling. By extracting the direct power sensing capability and replacing it with software-based measurement of utilization and frequency followed by power calculation, the system achieves real-time measurement without adding physical complexity to the GPU hardware.
Solution Approach 2:
The patent substitutes the mechanical/electrical approach of installing a physical power sensor in the GPU with a software-based approach using API calls to monitor GPU utilization and driving frequency. This replacement eliminates the need for additional hardware components and driver registrations, reducing device complexity while maintaining real-time measurement capability.
3Measurement precision
If only GPU power consumption is measured through power sensor registration, then measurement focus is improved, but overall power analysis becomes difficult
Solution Approach 1:
The patent applies universality by creating a power measurement approach that can be applied to the entire heterogeneous multi-core system, not just the GPU. The same power model and measurement methodology used for the GPU can be extended to measure CPU and other component power consumption, enabling comprehensive system-wide power analysis while maintaining the ability to focus on specific components when needed.
Data Source
AI summary
Provided is a method for measuring power of a graphics processing unit. The method includes changing a utilization of the graphics processing unit through an application programming interface (API), measuring and storing the utilization or a driving frequency of the graphics processing unit for each trace time, measuring and storing power consumption of the graphics processing unit for each trace time, and synchronizing the utilization of the graphics processing unit with the power consumption according to a stored trace time and calculating a power coefficient for each driving frequency with reference to synchronized information.