基於機器學習與考量動態壓降之電源網路可繞度最佳化

Abstract

在現代電路設計中，電源網路 (power/ground network) 是其中相當重要的部分，因為電源網路的配置將密切影響可靠性問題，如電壓降 (IR drop)、電遷移 (electromigration)，以及面積成本如金屬面積、訊號線繞線率 (signal net routability)。電源網路中的動態壓降 (dynamic IR drop) 是先進技術節點中最關鍵的問題之一。過度的動態壓降會降低電路性能，並導致潛在的功能性錯誤，但使用商用軟體模擬分析動態壓降值卻非常耗時。對於動態壓降的限制，大多業界作法都傾向於超規格設計 (overdesign) 電源網絡，但此做法卻會減少繞線資源 (routing resources) 並引起了繞線擁擠 (routing congestion)。而現有的基於機器學習 (machine learning) 的方法僅針對動態壓降預測，而沒有考慮受到電源網路影響的訊號線繞線率。 在本篇論文中，我們開發了一個兩階段的演算法流程，該流程包括用於訓練機器學習模型的數據預處理 (data preprocessing) 和多目標優化 (multi-objective optmization) 方案。在數據預處理階段，我們提出了一種有效提取周圍環境特徵的特徵工程 (feature engineering) 方法，這些特徵可用於同時預測動態壓降和繞線擁擠。在多目標優化方案中，我們採用前一階段的機器學習模型，並解決動態壓降與繞線資源之間的權衡問題。 實驗結果顯示，該演算法在模型精度和繞線資源優化方面均十分有效。我們的模型可以通過採用多任務學習 (multi-task learning) 以準確預測動態壓降和訊號線擁擠，從而獲得高達0.996的相關係數。實驗結果也顯示，通過採用的機器學習模型，該演算法能節省約10％的繞線資源，而不會顯著增加動態壓降峰值。與業界領先的商用軟體的模擬分析相比，我們的演算法還可實現高達48倍的顯著加速。

The power/ground (PG) network is an essential component in modern circuit designs, as the configuration of a PG network is closely related to the reliability issues (IR drop, electromigration) and area cost (metal area, signal net routability). The dynamic IR drop of a PG network is one of the most critical problems in an advanced technology node. Excessive IR drop slows down circuit performance and causes potential functional failures while obtaining the exact IR drop value is time-consuming by the simulation-based commercial tool. For the dynamic IR drop constraints, most industrial practices tend to over-design the PG network, reducing routing resources and incurring routing congestion. Existing machine-learning-based approaches target only dynamic IR drop prediction without considering the routability affected by the PG network. In this thesis, we develop a two-stage algorithm flow consisting of the data preprocessing for model training and the multi-objective optimization scheme. In the data preprocessing stage, we propose a feature engineering method that effectively extracts features containing neighboring information. These features can be used to predict dynamic IR drop and routing congestion simultaneously. In the multi-objective optimization scheme, we adopt the machine-learning model from the previous stage and solve the trade-off between dynamic IR drop and routing resources. Experimental results show that our algorithm is effective in both model accuracy and routing resources optimization. Our model can accurately predict dynamic IR drop and signal net congestion by adopting a multi-task learning scheme, achieving 0.996 high correlation coefficient. The experimental results also show that our algorithm can save about 10\% routing resources without worsening dynamic IR drop peak value by adopting the machine learning model. Our algorithm also achieves significant speedups of up to 48X, compared to the time-consuming dynamic IR drop simulation by a leading commercial tool.