A Generic Processor Temperature Estimation Method

Abstract

Most modern mobile embedded devices have the ability to increase their computational power typically at the cost of increased heat dissipation. This may result in temperatures above the design limit, especially if active cooling is inapplicable. Thus, it is necessary to consider processor temperature while scheduling tasks. This means estimating the change in temperature due to changed workload is crucial for high performance mobile embedded devices. To address this challenge, we first introduce a model to estimate the temperature and classify the system dependent model parameters. Then, to determine these parameters, we develop a new method, which can be applied on any mobile embedded device. The only requirement for our new method is learning the device characteristics by processing a certain task while recording the temperature with built-in sensors. Our results show that our method can achieve high accuracy within a short testing period. (( 25 th INTERNATIONAL WORKSHOP on Thermal Investigations of ICs and Systems )) September 2019, Lecco / IT www.therminic2019.eu ISBN 978-1-7281-2078-2 (( 25 th INTERNATIONAL WORKSHOP Thermal Investigations of ICs and Systems )) 2019 (( 25 th INTERNATIONAL WORKSHOP on Thermal Investigations of ICs and Systems ))

Improving Temperature Prediction Accuracy Using Kalman and Particle Filtering Methods

“…The thermal models in the literature [5][6][7][8] for computing the future temperature based on the current temperature and the rate of processing activities are mostly derived from the wellknown heat transfer equation:…”

“…The system temperature can be predicted by statistical methods, as described in [4], which use autoregressive moving averages (ARMA) to estimate future temperatures. On the other hand, more recent studies [5][6][7][8] combine statistical methods and the dynamic model of the system, thus combining past data and dynamic temperature behavior. In [5], the authors use a thermal model for such predictions.…”

“…[6] introduces a similar concept, however, only using temperature sensors. In [7], we introduced a thermal model and a method to find the system parameters. Our method only uses built-in temperature sensors and is applicable easily to each and every device since it does not require any additional hardware, such as a furnace or power sensor.…”

“…In this paper, we propose more accurate and robust future temperature prediction methods. Although our previously introduced method [7] can be applied periodically to adapt to environmental changes, we further improve it here, by extending the model first using the Kalman filtering method and then using the particle filtering method. In addition, we experimentally validate the new methods, that is, we compare (i) our original model, (ii) the model improved with a Kalman filter, and (iii) the model improved with a particle filter.…”

“…Our aim is to accurately predict the system temperature using a thermal model. In Section 2, we briefly explain the thermal model from [7] and the prediction method based on only this model. Then, we introduce two different methods to improve the prediction accuracy.…”