Autonomous Core Cluster Frequency Scaling With AMU Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional processor devices lack the ability to account for Quality-of-Service (QoS) requirements of executing workloads in making power management decisions, leading to inefficient frequency scaling.

Innovation Solution

A processor device autonomously manages core cluster frequencies using performance statistics, generating performance and energy-per-instruction models based on Activity Management Unit (AMU) data to identify a target frequency operating point, which is then adjusted by a Dynamic Voltage and Frequency Scaling (DVFS) circuit.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If conventional power management circuits use ACPI/CPPC to manage processor performance, then performance hints from the OS can enable better power efficiency, but the system cannot take into account the Quality-of-Service (QoS) requirements of executing workloads in making power management decisions

Engineering Contradiction:

Improvepower efficiencyVSAvoidQoS requirement consideration

Core Design Contradiction:

Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The system implements feedback by continuously monitoring AMU statistics from multiple frequency operating points and using this data to dynamically adjust frequency selections. The power management circuit collects performance statistics, generates performance models, and uses these models to make informed decisions about frequency scaling that balance power efficiency with QoS requirements.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system performs preliminary actions by pre-collecting AMU statistics across multiple frequency operating points and generating performance models in advance. This allows the power management circuit to have pre-computed performance data available when making real-time frequency decisions, enabling it to account for QoS requirements without compromising response time.

Inventive Principle:

Principle #10Preliminary action

2Power

If the power management circuit collects AMU statistics for frequency management, then it can make informed power decisions, but it lacks the capability to autonomously determine optimal frequency points considering both performance and energy consumption

Engineering Contradiction:

Improvepower management capabilityVSAvoidautonomous frequency optimization

Core Design Contradiction:

PowerVSExtent of automation

Solution Approach 1:

The power management circuit performs self-service by autonomously generating performance models from collected AMU statistics and using these models to automatically determine optimal frequency operating points. The system serves itself by having the same circuit that collects statistics also perform the analysis and decision-making, eliminating the need for external intervention and enabling autonomous optimization of both performance and power consumption.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system applies parameter changes by dynamically adjusting frequency operating points based on generated performance models. The power management circuit modifies frequency parameters in real-time according to workload conditions and QoS requirements, transitioning between different frequency states to optimize the balance between performance and power consumption.

Inventive Principle:

Principle #35Parameter changes

3Speed

If frequency scaling is performed without considering workload QoS requirements, then power management decisions can be made quickly, but the frequency scaling becomes inefficient and does not optimize for actual workload needs

Engineering Contradiction:

Improvefrequency scaling speedVSAvoidfrequency scaling efficiency

Core Design Contradiction:

SpeedVSProductivity

Solution Approach 1:

The system performs preliminary actions by pre-generating performance models from AMU statistics collected across multiple frequency operating points. This preparation work enables the power management circuit to make quick frequency scaling decisions later without having to perform complex analysis in real-time, thus maintaining fast response speed while ensuring efficient, QoS-aware frequency selection.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system uses feedback by continuously monitoring AMU statistics and using this information to adjust frequency scaling decisions. The performance models are generated based on actual workload behavior observed through AMU counters, creating a feedback loop that ensures frequency scaling is both fast and efficient by adapting to actual workload QoS requirements.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS12405651B2Autonomously managing core cluster frequencies using performance statistics in processor devices

Publication Date: 2025.09.02 QUALCOMM INC

AI summary

Autonomously managing core cluster frequencies using performance statistics in processor devices is disclosed herein. In some aspects, a cluster power management circuit of a processor device collects Activity Management Unit (AMU) statistics for multiple processor cores for each of one or more frequency operating points over a time interval. Based on the AMU statistics, the cluster power management circuit generates a performance model representing processor performance as a function of frequency, and uses the performance model and a power consumption measurement to generate an energy-per-instruction (EI) model representing energy per instruction as a function of frequency. The cluster power management circuit then generates an advantage model based on a first rate of change of the performance model as a function of frequency and a second rate of change of the EI model as a function of frequency, and identifies a target frequency operating point based on the advantage model.