NUMA Multi-Node Power Management with Cache Flush and Unicast Interrupts

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Efficient power management in multi-node computing systems is challenging due to the complexity of powering down and awakening processing nodes, which leads to increased latency and reduced performance, especially in systems with multiple power controllers and non-uniform memory access architectures.

Innovation Solution

Implementing a power management system that utilizes power controllers to determine idle states, flush caches, invalidate probe filters, and send unicast messages for interrupt handling, thereby minimizing unnecessary node awakenings and optimizing power states.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If each power controller is connected to each other power controller in the system, then communicating when to power down the system becomes easier, but scaling the system becomes difficult and routing connections increases signal area

Engineering Contradiction:
Improvepower down communicationVSAvoidsystem scaling
Core Design Contradiction:
Ease of operationVSDevice complexity

Solution Approach 1:

The patent introduces a fabric interface as an intermediary component that mediates communication between power controllers and other system components. Instead of direct point-to-point connections between all power controllers, the fabric interface serves as a central communication hub that routes messages between nodes, simplifying the connection topology and enabling easier system scaling.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Use of energy by moving object

If nodes are powered down to reduce power consumption, then energy savings are achieved, but latency increases when nodes need to be awakened soon afterward

Engineering Contradiction:
Improvepower consumptionVSAvoidnode awakening latency
Core Design Contradiction:
Use of energy by moving objectVSLoss of time

Solution Approach 1:

The patent implements preliminary actions before powering down nodes by flushing caches and invalidating probe filters in advance. This ensures that when a node is awakened, it can resume operations quickly without experiencing prolonged latency, as the cache coherence state is already prepared and validated before the power transition occurs.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements feedback mechanisms where power controllers monitor system activity and node utilization patterns. Based on this feedback, the system intelligently determines when to power down nodes and when to keep them active, balancing power consumption against potential awakening latency by learning from historical usage patterns.

Inventive Principle:
Principle #23Feedback

3Reliability

If caches are flushed before powering down nodes, then cache coherence is maintained, but power management complexity increases

Engineering Contradiction:
Improvecache coherenceVSAvoidpower management steps
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the cache flushing operation with the power down sequence by integrating these operations into a unified power management protocol. The fabric interface combines cache coherence maintenance tasks with power state transitions, so that what would otherwise be separate complex operations are coordinated as a single atomic sequence, reducing overall management complexity.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3729233B1Multi-node system low power management
Publication Date: 2025.08.06 ADVANCED MICRO DEVICES INC
  • EP3729233B1 patent drawingFigure 1
  • EP3729233B1 patent drawingFigure 2
  • EP3729233B1 patent drawingFigure 3

AI summary

Systems, apparatuses, and methods for performing efficient power management for a multi-node computing system are disclosed. A computing system including multiple nodes utilizes a non-uniform memory access (NUMA) architecture. A first node receives a broadcast probe from a second node. The first node spoofs a miss response for a powered down third node, which prevents the third node from waking up to respond to the broadcast probe. Prior to powering down, the third node flushed its probe filter and caches, and updated its system memory with the received dirty cache lines. The computing system includes a master node for storing interrupt priorities of the multiple cores in the computing system for arbitrated interrupts. The cores store indications of fixed interrupt identifiers for each core in the computing system. Arbitrated and fixed interrupts are handled by cores with point-to-point unicast messages, rather than broadcast messages.