ML-Driven SoC Power Adaptation via Telemetry
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing adaptive processors rely on expert-created rules and first-order statistical models for hardware adaptation, which are inadequate for large-scale commercial deployment due to high heuristic accuracy limitations and poor performance in capturing fine-grained workload predictions, especially with non-linear workload behaviors.
Innovation Solution
A machine learning-driven adaptation system that uses a telemetry system to gather metadata across a System on a Chip (SoC) and a microcontroller to execute machine learning models, enabling dynamic management of processor resources by disabling out-of-order components and throttling frontend operations based on predictive models, without requiring specialized hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If expert-created rules and first-order statistical models are used for hardware adaptation, then device complexity is reduced, but measurement precision and prediction accuracy deteriorate
Solution Approach 1:
The patent replaces traditional mechanical/heuristic control policies with machine learning models that can capture non-linear workload behaviors. The ML models process telemetry data to predict workload phases with high accuracy, enabling precise hardware adaptation decisions without relying on simple statistical models or expert rules.
Solution Approach 2:
The patent introduces machine learning models as an intermediary between raw telemetry data and hardware adaptation decisions. The ML models act as a bridge that transforms complex telemetry inputs into accurate workload phase predictions, which then guide hardware configuration changes with minimal complexity in the control logic itself.
2Measurement precision
If more data streams are considered for heuristic rules, then measurement precision improves, but device complexity increases significantly
Solution Approach 1:
The patent employs universal machine learning models that can process multiple data streams simultaneously through a single unified framework. The same ML model architecture handles various telemetry inputs (performance counters, power metrics, workload characteristics) without requiring separate control policies for each data stream, thus maintaining low device complexity while improving measurement precision.
3Measurement precision
If neural networks are used for fine-grained workload prediction, then measurement precision improves, but device complexity and hardware requirements increase
Solution Approach 1:
The patent implements dynamic hardware adaptation where the system continuously adjusts hardware configurations based on real-time workload phase predictions from ML models. The adaptation mechanism dynamically switches between different hardware states (e.g., enabling/disabling out-of-order execution, adjusting frontend throttling) according to predicted workload characteristics, achieving fine-grained control without permanent hardware modifications.
Solution Approach 2:
The patent changes operational parameters of the processor based on ML predictions, such as adjusting clock frequencies, enabling/disabling specific execution units, or modifying cache policies. These parameter changes are driven by workload phase predictions and allow fine-grained power management without requiring dedicated hardware for each adaptation scenario.
4Use of energy by moving object
If hardware is adapted tightly to workload phases, then power efficiency improves, but adaptability requirements and control complexity increase
Solution Approach 1:
The patent implements a feedback-driven adaptation system where telemetry data from the processor feeds into ML models that predict workload phases, which then trigger hardware configuration changes. The system continuously monitors performance and power metrics, adjusts hardware settings based on predictions, and uses the results to refine future predictions, creating a closed-loop feedback mechanism that improves power efficiency while managing adaptability requirements.
Data Source
AI summary
A system on a chip is described that comprises a processor and a set of memory components that store instructions, which when executed by the processor cause the system on a chip to: generate, by a set of data collectors of a telemetry subsystem, a set of streams of telemetry metadata describing operation of the processor, forward one or more streams of telemetry metadata from the set of streams of telemetry metadata to a set of machine learning-driven adaptation decision models, receive, from the set of machine learning-driven adaptation decision models, a set of configuration parameters for controlling operation of the processor based on the one or more streams of telemetry metadata, and modify operation of the processor based on the set of configuration parameters.