Stacked Die Thermal Management Using Virtual Victim-Die Sensing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing thermal management schemes in multichip modules fail to accurately predict temperature for victim dies due to neglecting thermal impacts from neighboring dies in horizontal and/or vertical stacks, leading to inaccurate performance management and potential reliability issues.
Innovation Solution
A thermal management scheme that utilizes telemetry information from local die sensors and neighboring dies to generate a virtual thermal sensing, allowing for accurate prediction of temperature for target dies or intellectual property domains, and enabling precise performance management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If thermal sensors are placed only on local dies without considering neighboring dies, then device complexity is reduced, but measurement precision of temperature for victim dies deteriorates
Solution Approach 1:
The patent introduces an intermediary computational model that processes temperature data from multiple aggressor dies to predict temperatures in victim dies. This mediator layer (the prediction model) enables accurate temperature measurement in victim dies without physically placing sensors there, thus maintaining low device complexity while achieving high measurement precision.
Solution Approach 2:
The patent creates a virtual copy of the thermal sensing capability by using machine learning models to predict temperatures in victim dies based on patterns learned from aggressor die temperature data. This virtual thermal sensor copy eliminates the need for physical sensors in victim dies while providing accurate temperature measurements.
2Measurement precision
If thermal sensors are placed on all dies including victim dies, then measurement precision of temperature is improved, but device complexity and manufacturing cost increase
Solution Approach 1:
The patent creates virtual thermal sensors through machine learning models that predict victim die temperatures based on aggressor die sensor data. This copying approach provides accurate temperature measurements without the need for additional physical sensors, thereby maintaining measurement precision while reducing device complexity and sensor quantity.
Solution Approach 2:
The prediction model serves as an intermediary that translates readily available aggressor die temperature data into accurate victim die temperature predictions, eliminating the need for direct physical measurement in victim dies and reducing overall sensor requirements.
3Ease of operation
If performance management is based on inaccurate temperature data from existing schemes, then ease of operation is maintained, but reliability of multichip modules deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the machine learning model continuously learns from actual temperature measurements and operational data to improve its predictions. This feedback loop ensures that performance management decisions are based on increasingly accurate temperature data, thereby improving reliability while maintaining ease of operation through automated model updates.
Solution Approach 2:
The prediction model acts as an intermediary layer between raw sensor data and performance management decisions, providing filtered and accurate temperature information for victim dies. This intermediary ensures reliable decision-making without complicating the operational workflow, as the model handles the complexity of multi-die thermal interactions automatically.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The virtual thermal sensing scheme decouples dielet design, reduces the number of thermal sensors needed, and enhances thermal management by accurately predicting temperatures and adjusting performance parameters such as frequency, power, and refresh rate.
Implementation Method 1
hot spots in a first set of dies (aggressor dies) can impact thermal constraints of a second set of dies (e.g., victim dies)
Implementation Method 2
thermal sensors near hotspots of their respective dies
Data Source
AI summary
A thermal management scheme, for a multichip module, that is aware of various dies in a stack (horizontal and/or vertical) and heat generated from them, local hot spots in a victim die, and hot spots in aggressor die(s). Each victim die receives telemetry information from thermal sensors located in aggressor dies as well as local thermal sensors in the victim die. The telemetry information is used to enable a virtual sensing scheme where temperature for a target die (e.g., a victim die) and/or its intellectual property (IP) domain is estimated or calculated. The estimated or calculated temperature is then used for performance management of the victim and/or aggressor dies in the stack.