ML Runtime Tuning of Processing Units for Power-Latency Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing power-performance management in computing devices is reactive and does not leverage learned behaviors, leading to inefficiencies and latency in processing unit operations.
Innovation Solution
A machine learning-based optimization model is trained to proactively adjust processing unit settings using activity data, incorporating metrics like utilization and memory usage, to balance performance and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If reactive power-performance management is used, then system responsiveness is maintained, but latency and inefficiency increase
Solution Approach 1:
The system performs preliminary actions by proactively adjusting processing unit settings before performance degradation occurs. The optimization model continuously monitors activity data and preemptively modifies operational parameters based on learned patterns, preventing latency issues rather than reacting to them after they occur.
Solution Approach 2:
The system implements continuous feedback loops where the optimization model monitors runtime activity data, evaluates performance metrics, and adjusts processing unit settings dynamically. This closed-loop feedback mechanism enables the system to maintain responsiveness while minimizing latency through data-driven decisions.
2Productivity
If processing unit settings are adjusted dynamically, then performance efficiency is improved, but system complexity increases
Solution Approach 1:
The optimization model serves as an intermediary layer between the processing units and the control system. It abstracts the complexity of dynamic settings adjustment by encapsulating the decision-making logic within the model, which processes activity data and generates optimized settings without requiring complex control circuitry or manual intervention.
Solution Approach 2:
The system achieves self-service through the optimization model that autonomously monitors runtime activity data, determines optimal settings, and adjusts processing unit parameters without external intervention. This self-managing capability improves performance efficiency while avoiding the complexity of external control mechanisms.
3Use of energy by moving object
If machine learning models are used for optimization, then power consumption is reduced, but computational overhead increases
Solution Approach 1:
The system applies partial action by using lightweight machine learning models that perform only the essential optimization functions needed. Rather than implementing comprehensive complex models, the system uses simplified models that provide sufficient optimization benefit while minimizing computational overhead and training requirements.
Data Source
AI summary
Disclosed are apparatuses, systems, and techniques that use machine learning techniques for determination and tuning of runtime settings of processing units. In one embodiment, a computing device, which includes one or more processing units, processes, using a machine learning model, a runtime activity data to generate settings for the processing unit(s). The runtime activity data characterizes an execution of a computing application on the processing unit(s). The computing device then modifies, using the generated settings, the execution of the computing application on the processing unit(s).