類神經網路於中央處理器風扇轉速最佳化與散熱片狀態偵測之應用
摘要
隨著科技的日新月異，中央處理器(CPU)的頻率與運算能力也越來越快，相對的，我們必須面對的是CPU在高速運算時所產生的高溫與散熱問題。一般而言，目前市面上常見的CPU散熱方式有以下兩種:(以散熱器的型態來區分)(1)主動式散熱器: 由一個脈波寬度調變(PWM)風扇與散熱片(Heat Sink)組成，並由風扇轉速來控制流過散熱片的氣流強弱;(2)被動式散熱器: 體積較大且需搭配主機機殼內獨立之風扇與導風罩一起使用，方可達到散熱的效果。由於成本與便利性的考量，一般個人或套裝PC皆是以主動式散熱器為主;被動式散熱器則是多用於伺服器等具多個發熱源的系統。 由於主動式散熱器是由一個PWM風扇與散熱片所組成，為了找出CPU最佳的散熱參數(散熱片材質、結構與風扇轉速等)，一般主機板廠商在產品設計階段，皆需投入大量人力、時間與各種專業的設備，進行多次的實驗，以保證產品的散熱效能可以符合CPU的規範。然而，目前市面上的個人電腦(PC)或套裝電腦，為了支援市面上種類繁複的CPU並縮短產品上市的時間，主機板廠商通常會以能支援的CPU中功率最大者為其散熱方案的設計目標，如此雖然可以保證產品的散熱效能足以符合各種CPU的規範，但，如果使用者使用的是功率較低的CPU時，則在散熱片與風扇轉速不變的狀況下，將會產生Over-Cooling的問題，造成能源不必要的浪費與較高的噪音值。 本論文提出一個CPU風扇轉速最佳化與散熱片狀況偵測的方法，利用類神經網路之放射狀基底函數網路來建立電腦內CPU溫度的模型，並利用此一模型搭配CPU的Thermal Profile，來離線找出另一組可令CPU溫度維持在規範範圍內的風扇轉速資料，並利用此一資料來訓練另一個類神經網路控制器，從而實做一個CPU風扇轉速控制軟體。而原本用以建立電腦內CPU溫度模型的類神經網路，則可用來做為散熱器狀態偵測的用途，讓使用者無需打開機殼便可確認散熱器的狀態。 經實驗證明，兩顆不同散熱設計功率(TDP)的CPU在兩部不同品牌的系統上，以本論文實作之CPU風扇轉速控制軟體來控制CPU風扇，可以比系統上原有的風扇轉速控制機制，平均降低18.65% (ASUS+E4600)、56.03% (ASUS+X3220)、45.81% (GIGABYTE+E4600)、1.81% (GIGABYTE+X3220)的風扇轉速(風扇能源消耗與風扇噪音值也相對地下降)。而散熱器狀態偵測在不加遮罩的狀況下平均準確度可達98.20% (ASUS+E4600)、94.71% (ASUS+X3220)、98.24% (GIGABYTE+E4600)、98.40% (GIGABYTE+ X3220)。

關鍵字：主動式散熱器、脈波寬度調變、風扇轉速控制、Over-Cooling、放射狀基底函數網路、Thermal Profile、TDP、散熱器狀態偵測

Application of Neural Networks in CPU Fan Speed Optimization and Heat Sink Status Detection
Abstract
With technology’s revolution and innovation, CPU frequency and computing capability become more and more higher and faster. But, at same time, we also have to face the problem of high temperature and heat dissipation that generated by CPU. Generally, there are two common solutions for CPU heat dissipation (based on the types of heatsinks): (1) Active Heatsinks: Consists of one PWM fan and one heatsinks, and control the airflow of heatsinks via variable PWM fan speed; (2) Passive Heatsinks: Larger than Active Heatsinks, and require in-depth knowledge of the airflow and air-ducting techniques to manage the airflow in the chassis. Consider cost and convenience, usually, Active Heatsinks solution will be adopted on the most PC or own-brand computer, and Passive Heatsink is mainly used on the system (ex. Server) with multi-heat sources. Since Active Heatsinks consist of two parts - PWM fan and heatsinks, motherboard vendors usually have to spend lots of man-power、time and special equipments in their experiments for finding out the best thermal solution (like material and structure of heatsinks, corresponding fan speed…etc) during product design phrase and make sure the performance of that thermal solution can meet CPU thermal specifications. However, in order to support all CPU(s) that in CPU support list and shorten the lead-time, the mother board vendors often adopt a strategy - design the thermal solution of their products to fulfill the CPU with maximum power in the CPU support list. This strategy guarantees their products are able to meet every CPU’s thermal specifications always but it could cause Over-Cooling problem. Unnecessary power consumption and higher noises could be caused. In this thesis, the methods of CPU Fan Speed Optimization and Heat Sink Status Detection will be presented. Two RBFN (Radial Basis Function Networks) will be used. The first RBFN (Radial Basis Function Networks) will be used for modeling a CPU thermal model with the CPU thermal data we collect. And, a set of data for training second RBFN will be generated by using the combination of the first RBFN and CPU Thermal Profile. After second RBFN training finish, it will be integrated into a CPU fan speed control application program and used to control CPU fan speed continually and smartly under OS (operating system). And, the first RBFN will be used for detecting the CPU heatsinks status. End-user will not need to open the chassis for checking the CPU heatsinks status anymore. To demonstrate the performance of the proposed methods, two CPU(s) with different TDP have been tested on two different systems with RBFN fan speed controller and original thermal solutions separately. And, the test result shows the RBFN fan speed controller can reduce 18.65% (ASUS+E4600)、56.03% (ASUS+X3220)、45.81% (GIGABYTE+E4600)、1.81% (GIGABYTE+X3220) fan speed (on average) than original thermal solutions(fan power consumption and fan noise can be lowered too) . And, the average predictability of RBFN Heat Sink Status Detector are 98.20% (ASUS+E4600)、94.71% (ASUS+X3220)、98.24% (GIGABYTE+E4600)、98.40% (GIGABYTE+X3220) without fan mask.

Keywords : Active Heatsinks、PWM、Fan Speed Control、Over-Cooling、RBFN、Thermal Profile、TDP、Heatsinks Status Detection