Dynamic Accelerator Power Management via Workload Metadata Scheduling

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current solutions for dynamic power management in Multi-Chip Package (MCP) solutions with integrated accelerators are inadequate, as they fail to efficiently manage power consumption and reconfiguration, leading to unnecessary energy usage and suboptimal performance due to reliance on hardware autonomous or OS-driven methods that do not effectively eliminate idle power consumption and require costly reset cycles.

Innovation Solution

Implementing a scheduler that dynamically controls accelerator devices by reading workload metadata to enable or disable power based on resource requirements, using a management controller and operating system to offload processes, and performing hot-plug operations, thereby optimizing power usage and avoiding costly reset cycles.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If accelerators are integrated into traditional processor die to create MCP solutions, then processing capability is improved, but power consumption increases when accelerators are idle

Engineering Contradiction:

Improveprocessing capabilityVSAvoidpower consumption

Core Design Contradiction:

PowerVSUse of energy by moving object

Solution Approach 1:

The system dynamically enables or disables accelerator devices based on real-time workload demands. The scheduler monitors workload metadata and adjusts the operational state of accelerators accordingly, transitioning them between active and idle states to match actual processing needs, thereby reducing unnecessary power consumption while maintaining processing capability when needed.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes the operational parameters of accelerator devices by enabling or disabling them based on workload requirements. This parameter change allows the system to adapt power consumption levels according to actual processing demands, resolving the contradiction between maintaining processing capability and reducing idle power usage.

Inventive Principle:

Principle #35Parameter changes

2Use of energy by moving object

If accelerators are disabled to reduce power consumption, then power efficiency is improved, but reconfiguration requires costly reset cycles

Engineering Contradiction:

Improvepower efficiencyVSAvoidreset cycle time

Core Design Contradiction:

Use of energy by moving objectVSLoss of time

Solution Approach 1:

The scheduler performs preliminary actions by monitoring workload metadata and proactively managing accelerator states before workload changes occur. This allows smooth transitions between active and idle states without requiring costly reset cycles, as the system prepares for state changes in advance based on predicted workload demands.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The scheduler acts as an intermediary between the workload requirements and the accelerator devices. It manages the transition states and coordinates the enabling/disabling of accelerators, preventing the need for costly reset cycles by mediating the reconfiguration process through intelligent workload analysis and gradual state transitions.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Extent of automation

If hardware autonomous or OS-driven power management is used, then automation is improved, but idle power consumption is not effectively eliminated

Engineering Contradiction:

Improvepower management automationVSAvoididle power consumption

Core Design Contradiction:

Extent of automationVSUse of energy by moving object

Solution Approach 1:

The system implements feedback mechanisms where the scheduler continuously monitors workload metadata and accelerator performance metrics. This feedback loop enables the system to make informed decisions about accelerator states, adjusting power management based on actual workload demands rather than relying solely on hardware autonomous or basic OS-driven methods, thereby effectively eliminating idle power consumption.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The scheduler enables accelerators to serve themselves by dynamically adjusting their operational states based on workload requirements. The system self-manages the enabling and disabling of accelerator devices without requiring external intervention, optimizing power consumption while maintaining processing capability through autonomous decision-making based on workload metadata.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS11157064B2Techniques to dynamically enable and disable accelerator devices in compute environments

Publication Date: 2021.10.26 INTEL CORP

AI summary

Various embodiments are generally directed to an apparatus, method and other techniques to send a power operation initiation indication to the accelerator device via the subset of the plurality of interconnects, the power operation initiation indication to indicate a power operation to be performed on one or more infrastructure devices, receive a response the accelerator device, the response to indicate to the processor that the accelerator is ready for the power operation, and ucause the power operation to be performed on the accelerator device, the power operation to enable or disable power for the one or more of the infrastructure devices.