對於關鍵時序路徑上壓降熱點之準確向量動態壓降預測

Abstract

在關鍵時序路徑上過大的電路壓降會導致時序違規，因此電路壓降簽核至關重要。現有的機器學習方法在預測向量的動態電路壓降時，對於關鍵時序路徑上的壓降熱點表現出較大的預測誤差。本文提出了一種精確預測向量的動態電路壓降的方法，並且專注於預測關鍵時序路徑上的壓降熱點。我們的方法包括提取關鍵時序元件、創建電阻加權特徵，以及引入系統性的特徵選擇流程以提高預測準確度。實驗結果顯示，與以往的研究相比，我們的方法在關鍵時序路徑上的壓降熱點的平均絕對誤差減少了45.7%，最大誤差則減少了40.2%。我們的研究為關鍵時序路徑上的壓降熱點提供精確的動態電路壓降預測，以進一步輔助時序簽核過程。

IR-drop sign-off is very important since excessive IR-drop on critical timing paths can lead to timing violations. Existing machine learning methods to predict vector-based dynamic IR-drop exhibit large prediction errors for IR-hotspots on critical timing paths. This paper presents an approach for accurate vector-based dynamic IR-drop prediction focusing on IR-hotspots on critical timing paths. Our method includes the extraction of critical timing cells, creation of resistance-weighted features, and introduction of a systematic feature selection flow to enhance prediction accuracy. Experimental results show an average reduction of 45.7% in mean absolute error (MAE) and 40.2% in maximum error (MaxE) for IR-hotspots on critical timing paths compared to previous works. This research aids the timing sign-off process by offering accurate dynamic IR-drop prediction for IR-hotspots on critical timing paths.