Data Annealing for xPU Power Estimation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methodologies for building accurate and real-time power estimation models for computing processors (xPU) and their associated system units are unstable and unreliable under various physical and environmental factors, making them unsuitable for high-performance computing environments.
Innovation Solution
A data annealing process is used to purify raw electrical and operational state-related data, followed by a machine learning-based process to build a characteristic model that provides robust power and performance estimation across all xPU variants, both pre-silicon and post-silicon stages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If existing methodologies for power estimation modeling are used, then additional circuits and extensive data processing can be employed, but the model stability and reliability deteriorate under various physical effects and environmental factors
Solution Approach 1:
The patent applies parameter changes by transforming the data representation through coordinate transformations and feature engineering. The raw measurement data is converted into purified features through mathematical transformations that make the model invariant to certain physical effects and environmental variations, thereby improving reliability while maintaining manageable complexity
Solution Approach 2:
The patent introduces an intermediary data purification process that acts as a mediator between raw measurements and the final model. This intermediary layer processes and cleans the data by removing artifacts and normalizing features, creating a bridge that reduces the direct impact of environmental factors on model reliability
2Loss of time
If raw electrical data is used directly for modeling, then data processing time is reduced, but measurement precision deteriorates due to noise and artifacts
Solution Approach 1:
The patent applies preliminary action by performing data purification and feature extraction before the actual modeling process. The data is pre-processed to remove noise, artifacts, and irrelevant information, and transformed into meaningful features. This preliminary processing ensures high measurement precision is achieved efficiently without requiring extensive processing during the modeling phase
Solution Approach 2:
The patent extracts only the essential features and information from the raw electrical data through careful feature selection and extraction. By identifying and isolating the critical parameters that truly represent xPU characteristics, the system achieves high measurement precision while minimizing unnecessary data processing time
3Measurement precision
If extensive data processing is performed to clean raw data, then measurement precision is improved, but productivity deteriorates due to increased processing requirements
Solution Approach 1:
The patent transforms the data through parameter changes that simplify the processing requirements. By reparameterizing the data into meaningful features and characteristics, the system achieves high measurement precision with reduced processing complexity, thereby maintaining productivity while improving data quality
Solution Approach 2:
The patent creates a simplified representation or copy of the essential data characteristics through feature extraction. Instead of processing the entire raw dataset extensively, the system extracts and works with the critical features, achieving the same measurement precision with significantly improved productivity
Data Source
AI summary
A method of building a characteristic model includes: acquiring raw electrical data from a measurement system outside one or more processing units; acquiring operational state-related data from an information collector inside the one or more processing units; performing a data annealing process on the raw electrical data and the operational state-related data to obtain and purified electrical data and purified operational state-related data; and performing a machine learning (ML)-based process to build the characteristic model based on the purified electrical data and the purified operational state-related data.