An Effective and Accurate Data-Driven Approach for Thermal Simulation of CPUs

A data-driven reduced-order approach is applied to develop a thermal simulation model for a quad-core CPU, AMD ATHLON II X4 610e. The model is based on proper orthogonal decomposition (POD) that projects the physical domain of the CPU onto a functional space represented by a small set of basis funct...

Full description

Saved in:
Bibliographic Details
Main Authors: Jiang, Lin, Liu, Yu, Cheng, Ming-C.
Format: Conference Proceeding
Language:English
Subjects:
Computational modeling
data-driven learning
FEniCS
HotSpot
Predictive models
proper orthogonal decomposition
Thermal conductivity
Thermal decomposition
Thermal simulation of CPUs
Thermomechanical processes
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:A data-driven reduced-order approach is applied to develop a thermal simulation model for a quad-core CPU, AMD ATHLON II X4 610e. The model is based on proper orthogonal decomposition (POD) that projects the physical domain of the CPU onto a functional space represented by a small set of basis functions (or modes). To generate an optimal set of modes, these modes are trained by thermal solution data collected from numerical simulation tools, FEniCS and HotSpot Grid. If good-quality data are used in the training process, the process optimizes the POD modes that are then able to offer a very accurate simulation model with a very small numerical degree of freedom (DoF). Each of the developed POD models is verified against its simulation tool used in its training. A very accurate prediction is observed in the POD model derived from FEniCS with a reduction in the numerical DoF by nearly 5 orders of magnitude, which amounts to more than a 3-order reduction in computing time. The POD model derived from HotSpot Grid is however not able to offer accurate simulation due to its inadequate data quality.
ISSN:2694-2135
DOI:10.1109/ITherm51669.2021.9503183