標題: 統合性的最佳化理論與架構在電子設計自動化應用之研究
A Unified Optimization Framework for Electronic Design Automation
作者: 余紹銘
Shao-Ming Yu
李毅郎
李義明
Yih-Lang Li
Yiming Li
資訊科學與工程研究所
關鍵字: 最佳化;平台;電子設計;optimization;framework;electronic design
公開日期: 2007
摘要: 本研究之目的在於建立一個統合性的最佳化理論與架構,並且應用於電子設計自動化之領域。電子產業中有許多商用電腦輔助工具(CAD Tools),此類工具多著重於模擬實務甚至提供客製化的功能,可應用於當前技術設計。在前瞻的理論發展與技術開發,研究學者往往需要重新撰寫最佳化暨模擬程式,但這過程是非常複雜且耗時。因此如何建立一個最佳化架構,能夠很方便地將研究議題導向成一個最佳化問題,且能整合不同的電腦模擬軟體與最佳化方法來協助前瞻技術的開發是一項重要的課題。 本研究基於上述概念,希望能開發出一套具備高適用性、高性能的統合性最佳化架構。此研究將藉由整合各種不同的生物演化觀點之最佳化技術、數值最佳化方法與C++程式語言之物件導向設計,而建立了一個統合性的最佳化架構,使之可以適用於處理工程最佳化問題,特別是電子資訊產業的問題。在此架構中,問題定義與最佳化理論方法將會是各自獨立的兩個部分,因此使用者只要藉由程式介面定義並撰寫他們的問題,即可使用已建立之最佳化方法來求解問題。同時使用者也有高自由度可以任意添加新的最佳化方法到此架構中,以達到較佳的延伸性。 在不考慮數學上的收斂特性下,本研究也同時提出了一個混和式的最佳化方法。在此混和式的方法中,先藉由生物演化觀點演算法在問題的全域範圍內尋求一個解答,再利用數值最佳化方法將此全域區間的解答做進一步的改善,之後將數值最佳化方法所找出的區域空間內最佳解,重新送回生物演化觀點演算法中繼續求解。在各類電子資訊產業所遭遇的問題都是非常複雜的,而且通常並不能確定是否每個問題都有最佳解,因此對於業界而言只需要取得一個能滿足所有設計條件的解答即可。在本研究所提出的混和式方法中,藉由數值最佳化方法與生物演化觀點演算法的不斷交替使用,以及引入了各類電子資訊領域中的專業知識,可以針對複雜的電子設計問題,自動尋求出一個符合設計需求的解答。 在本研究中,進一步運用此概念整合不同的電腦輔助工具與自行研究的模擬程式,並且已經成功的應用在一個65 奈米的互補式金氧半電晶體(Complementary Metal Oxide Semiconductor;CMOS)的製程參數設計問題上。在此應用中,藉由與一個自行研究開發的半導體隨機摻雜濃度擾動效應分析程式以及一些工程經驗的結合,可以針對一個CMOS電晶體的設計規格,找出一組符合需求的製程參數組,此參數組同時也能抑制電晶體製造中因為隨機摻雜濃度所造成的電特性擾動現象。藉由與實際實驗數據的比對結果,驗證了此方法的準確性與可行性。此概念也同時被應用於半導體參數萃取、超大型積體電路設計以及通訊系統中的天線設計最佳化等問題上,而且也都得到了良好的結果。此統合性最佳化架構之程式碼已經提供在公開的網路上(http://140.113.87.143/ymlab/uof/)。
In the modern microelectronics industry, there are some kinds of computer-aided design tools (CAD tools) to assist engineers complete simulation jobs which can verify and estimate the performance of their designs. However, to satisfy the design targets, engineers must base on the simulation result to adjust the design parameters, and again feed the adjusted parameters to retrieve the improved result. Currently, such routine work mostly performed by engineers with expertise. Therefore, a well defined optimization platform can assist engineers to solve problems more efficiently. This dissertation presents an object-oriented unified optimization framework (UOF) for general problem optimization. Based on biological inspired techniques, numerical deterministic methods, and C++ objective design, the UOF itself has significant potential to perform optimization operations on various problems. The UOF provides basic interfaces to define a general problem and generic solver, enabling these two different research fields to be bridged. The components in the UOF can be divided into problem and solver parts. These two parts work independently, allowing high-level code to be reused, and rapidly adapted to new problems and solvers. Without considering the mathematical convergence property, one hybrid intelligent technique for electronic design automation is also proposed and implemented in the UOF. In the proposed hybrid approach, an evolutionary method, such as genetic algorithm (GA), firstly searches the entire problem space to get a set of roughly estimated solutions. The numerical method, such as Levenberg-Marquardt (LM) method, then performs a local optima search and sets the local optima as the suggested values for the GA to perform further optimizations. The electronic design problems from the industry are very complicated and not always guaranteed to have an optimal solution. Therefore, the designers or engineers only need to find one suitable solution which can meet all specifications. By integrating with empirical knowledge, the proposed hybrid approach can automatically search solutions to match the specified targets in the electronic design problems. The purpose of the UOF is to assist the electronic design automation with various CAD tools. One application in 65nm CMOS device fabrication has been investigated. Integration of device and process simulation is implemented to evaluate device performances, where the developed approach enables us to extract optimal recipes which are subject to targeted device specification. Fluctuation of electrical characteristics is simultaneously considered and minimized in the optimization procedure. Compared with realistic fabricated and measured data, this approach can achieve the device characteristics, and can reduce the threshold voltage fluctuation at the same time. Other applications including device model parameter extraction, very large scale integration (VLSI) circuit design and the communication system antenna design are also implemented with the UOF and presented in this dissertation. The results confirm that UOF has excellent flexibility and extensibility to solve these problems successfully. The developed open-source project is available in the public domain (http://140.113.87.143/ymlab/uof/).
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009323802
http://hdl.handle.net/11536/79156
Appears in Collections:Thesis


Files in This Item:

  1. 380201.pdf