Heterogeneous Processing Unit Frequency Control via Sensitivity Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional power management techniques for heterogeneous processing units, such as CPUs and GPUs, often break down in applications requiring cooperative execution due to performance and thermal coupling, leading to suboptimal performance and increased power consumption.
Innovation Solution
A performance controller evaluates frequency sensitivities across CPU and GPU cores using activity metrics to dynamically adjust their frequencies, balancing performance and power consumption by shifting power resources based on workload dependencies and thermal constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the frequencies of CPU or GPU cores are boosted to improve performance, then performance is improved for single-component workloads, but performance coupling and thermal coupling cause the boost algorithms to break down in cooperative execution applications
Solution Approach 1:
The patent merges the frequency control of CPU and GPU cores into a unified approach by introducing a shared frequency sensitivity metric that considers both components simultaneously. The performance controller evaluates frequency sensitivities of both CPU and GPU cores together, adjusting their frequencies in coordination rather than independently, thereby resolving the performance coupling issue while maintaining thermal headroom.
Solution Approach 2:
The patent implements feedback mechanisms by continuously monitoring activity metrics from both CPU and GPU cores to dynamically adjust their frequency states. The performance controller uses this feedback to evaluate frequency sensitivities and make real-time decisions about power allocation, ensuring that frequency boosts are applied only when they improve overall system performance while respecting thermal constraints.
2Productivity
If thermal headroom is used to increase the frequency of active processing units, then performance is improved, but thermal coupling causes premature throttling of GPU cores
Solution Approach 1:
The patent applies dynamic frequency adjustment by continuously evaluating frequency sensitivities and activity metrics to determine optimal frequency states for both CPU and GPU cores. The performance controller dynamically allocates power resources based on real-time workload characteristics and thermal conditions, allowing flexible frequency modulation that adapts to changing system states without causing thermal throttling.
Solution Approach 2:
The patent changes operational parameters by introducing a unified frequency sensitivity metric that transforms thermal management from a static constraint into a dynamic optimization parameter. The performance controller uses this metric to adjust frequency and power allocation in a coordinated manner, enabling the system to operate closer to thermal limits without premature throttling while maintaining stable temperatures through balanced power distribution.
3Productivity
If power is allocated to boost CPU frequency for control-divergent applications, then performance is improved, but power coupling causes suboptimal performance in graphics-intensive applications
Solution Approach 1:
The patent applies local quality by evaluating the specific workload characteristics and frequency sensitivities of each processing component independently before allocating power. The performance controller determines whether to boost CPU or GPU frequency based on the actual application type (control-divergent vs. graphics-intensive), directing power resources to the component that will provide the greatest performance benefit for that specific workload, thereby avoiding wasteful power consumption.
Solution Approach 2:
The patent changes power allocation parameters by introducing a unified frequency sensitivity metric that enables intelligent power distribution decisions. The performance controller uses this metric to dynamically adjust power allocation between CPU and GPU based on real-time activity metrics and frequency sensitivities, ensuring that power is directed to the component most needed for the current workload while avoiding unnecessary power consumption in the other component.
Data Source
AI summary
A method includes controlling active frequency states of a plurality of heterogeneous processing units based on frequency sensitivity metrics indicating performance coupling between different types of processing units in the plurality of heterogeneous processing units. A processor includes a plurality of heterogeneous processing units and a performance controller to control active frequency states of the plurality of heterogeneous processing units based on frequency sensitivity metrics indicating performance coupling between different types of processing units in the plurality of heterogeneous processing units. The active frequency state of a first type of processing unit in the plurality of heterogeneous processing units is controlled based on a first activity metric associated with a first type of processing unit and a second activity metric associated with a second type of processing unit.