針對時序工程修改命令的快速準確動態電壓降預測

Abstract

時序（timing）工程修改命令 (engineering change order, ECO）可能導致新的電壓降（IR-drop）熱點（hotspot），這些熱點需要在簽發（signoff）前通過額外的電壓降工程修改命令進行修復。在這篇論文中，我們提出了一種基於機器學習的做法，能預測時脈樹（clock tree）工程修改命令後，電路中每個單元（cell）的動態電壓降，以便讓我們找出時脈樹工程修改命令導致的新電壓降熱點。在機器學習模型訓練這方面，我們結合了時脈樹工程修改命令前後設計的特徵，並提出了四種新特徵：目標單元差異特徵、鄰近單元框形平均特徵、鄰近單元框形分佈特徵，和鄰近單元暫態功率特徵。我們的實驗結果顯示，相比於運用坊間軟體進行電壓降分析，我們的作法加速了至少九倍。預測結果的平均絕對百分比誤差也小於3%。我們的方法還可以在大多數情況下，有效找出時脈樹工程修改命令後的電壓降熱點。

Timing engineering change order (ECO) may result in new IR-drop hotspots, which are required to be fixed by additional efforts in IR-drop ECO before signoff. In this thesis, we proposed a machine-learning-based method to predict the dynamic IR-drop of every cell in a clock tree ECO revised design so that we can determine the new IR-drop hotspots resulting from the clock tree ECO. To train the machine learning model, we combine the features of a design before and after clock tree ECO, as well as propose four types of new features: target cell delta features, neighbor cell frame averaged features, neighbor cell frame distribution features, and neighbor cell transient power features. Our experimental results show that we achieved at least a 9x speed-up ratio compared to IR-drop analysis by commercial tools, with mean absolute percentage errors less than 3%. Our method can also improve the effectiveness of IR-drop hotspot identification after clock tree ECO in most cases.