Thermal Management of Processors by Using Temperature Prediction and Machine Learning

نوع: Type: thesis

مقطع: Segment: masters

عنوان: Title: Thermal Management of Processors by Using Temperature Prediction and Machine Learning

ارائه دهنده: Provider: Danial Haji aqa babaei

اساتید راهنما: Supervisors: Dr. Abbas Ramezani

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Dr. Moharram Mansourizadeh - Dr. Mehdi Abbasi

زمان و تاریخ ارائه: Time and date of presentation: 5 - 2 - 2023

مکان ارائه: Place of presentation: Amphitheater

چکیده: Abstract: Today, with the advancement of chip manufacturing technology, the power density is also increasing. The increase in power consumption density has led to an increase in temperature, which has a negative effect on efficiency, reliability, cost, lifetime of components and power due to leakage current. To solve the problem of increasing the processor temperature, in general, two static and dynamic methods can be used. The static method can be used during the design of chips and the dynamic method during the execution of programs by the chip. Various dynamic temperature management techniques are presented using the methods previously performed and temperature estimation along with task scheduling and migration, voltage and frequency regulation, thread injection or idle cycles. In dynamic management, various methods are used in different fields, for example, meta-heuristic methods, genetic algorithms and machine learning. Finally, unlike the static approach, in dynamic methods, the temperature is predicted and controlled before reaching the threshold limit using a temperature model. In this research, a model based on the combination of CNN and LSTM networks is proposed. To train the model, using the temperature sensors and performance counters inside the processor, a data set containing a suitable variety of temperature changes has been collected, and to collect this data set, a dataset based on the CPUSPEC2006 program has been used. To increase the accuracy of the model, other data with the names of historical features and control features have been extracted from the existing features. In the end, the proposed model for temperature prediction for time intervals of 2 to 5 seconds has been evaluated in different conditions. The results show that by choosing these features, the temperature of the next two seconds is predicted with an average absolute value error of less than one degree Celsius

فایل: ّFile: Download فایل