Digital Temperature Sensor Error Correction Using Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Digital temperature sensors in integrated circuit chips face errors due to fabrication process variations, environmental factors, and noise, which affect the accuracy of temperature measurements, particularly in advanced node chips and continuous running applications like AI servers.
Innovation Solution
The implementation of machine learning techniques using E-test data and intra-die variation data to learn and predict digital temperature sensor errors, enabling real-time error correction without the need for High Volume Manufacturing calibration, by training a model with absolute readings and electrical parameters to adjust sensor readings and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional digital temperature sensors are used without calibration, then manufacturing is simpler, but measurement precision deteriorates with errors of ±5°C
Solution Approach 1:
The patent applies preliminary action by training the machine learning model during the High Volume Manufacturing process to predict and correct temperature sensor errors. The model is trained in advance using E-test data and intra-die variation data, enabling automatic error correction without requiring post-manufacturing calibration for each sensor. This resolves the contradiction by achieving high measurement precision through preliminary model training while avoiding complex individual calibration procedures.
2Reliability
If thermal guard bands are implemented to ensure temperature safety, then reliability improves, but chip performance deteriorates due to reduced clock frequency
Solution Approach 1:
The patent applies feedback by using machine learning to continuously predict temperature sensor errors and provide real-time corrections to the temperature readings. The system feeds the predicted errors back into the temperature measurement process, enabling more accurate thermal monitoring. This allows the chip to operate closer to thermal limits with higher confidence, improving productivity by reducing unnecessary performance throttling while maintaining reliability through accurate temperature awareness.
3Measurement precision
If High Volume Manufacturing calibration is performed to reduce sensor errors, then measurement precision improves, but manufacturing complexity and cost increase
Solution Approach 1:
The patent applies self-service by enabling the temperature sensors to correct their own errors through the machine learning model. The model uses E-test data and intra-die variation data to predict errors specific to each sensor, allowing the sensors to self-correct without requiring external calibration procedures. This resolves the contradiction by achieving high measurement precision through self-service error correction while maintaining ease of manufacture by eliminating complex calibration steps.
Data Source
AI summary
Systems, apparatuses and methods may provide for chip technology including a memory structure having stored weights associated with a machine learning (ML) model, a plurality of digital temperature sensors to generate readings, and a classification engine to retrieve the stored weights from the memory structure and adjust the readings from the plurality of digital temperature sensors based on the weights and electrical parameters associated with the chip.