使用敏捷式基因探勘與隨機最佳化來改善類比電路合成的效率

Abstract

此篇論文提出了一個在類比電路的階層合成框架上的性能探勘技術和一個動態非均勻的模擬技巧。不同於規格針對性的設計，這篇研究主要是透過平行化的基因演算法探勘類比電路效能的極限，以達到尋找出比人為所不易找到的類比電路設計結果。不同於其他基於演化的拓撲探勘，這個方法能夠把性能視為基因組合來用在演化上且利用多人口群的特點來解決多目標的問題。在人口群裡所選擇的性能能夠利用重新針對的技巧轉換為器件參數。基於把器件參數正規化，一個概率動態模擬顯著地減少找到電路性能全域最佳解的收斂時間。這個演算法發展於分散式的OpenMp。實驗結果顯示出我們提出的類比電路合成方法在不同製程上的RFDA 和Op-Amp 電路能夠得到更好的運行時間且有更高的品質。

This thesis presents a performance exploration technique and a stochastic non-uniform simulation in hierarchical synthesis framework for analog circuit. Dierentfrom spec targeted designs, this proposed approach can help to search the solutions better than designers' expectation. A parallel genetic algorithm method is employed for performance exploration. Unlike other evolution-based topology explorations, this is a method that regards performance constraints as input genome for evolution and resolves the multiple-objective problem with the multiple-population feature. Populations of selected performance are transfered to device variables by re-targeting technique. Based on a normalization of device variable distribution, a probabilistic stochastic simulation signi?cantly reduces the convergence time to find the global optima of circuit performance. This algorithm is developed and run on distributed OpenMP. Experimental results show that our approach on radio-frequency distributed ampli?er (RFDA) and folded cascode operational amplier (Op-Amp) in di?erent technologies can obtain better runtime and higher quality in analog synthesis.