在現代IC設計和製造過程中超大型積體電路測試 (VLSI Testing) 擔任維持產品可靠度的重要角色。在常規的測試技術難以用合理的測試成本達到期望的測試覆蓋率 (Defect Coverage, DC) 的背景下，研究者應用資料分析技術開發出能夠進一步提高測試覆蓋率的測試方法。其中一種受到廣泛應用的方法稱為處於缺陷晶片群集中之正常晶片 (Good-Die-in-Bad-Neighborhood, GDBN) 檢測，此方法通過分析缺陷晶片群集的分佈樣態識別出具有高潛在風險的晶片，以滿足如車用電子、航太電子等對產品可靠度要求特別嚴苛的領域的需求。 近年來的研究應用深度神經網絡 (Deep Neural Network, DNN) 以追求更高的檢測準確度，然而深度神經網絡固有的高計算複雜度使得GDBN方法的檢測效率大幅降低。為了解決檢測效率退化問題，本文提出了一個高度平行的神經網絡架構名為FastGDBN，將推理時間複雜度降低至只與晶圓數量呈線性，並且藉由端到端 (end-to-end) 的架構減少了CPU與GPU之間的資料傳輸開銷 (CPU-GPU transfer overhead) 以及消除來自同一晶圓的多個晶片的檢測過程中的計算冗餘 (Redundant computation)。在 WM-811K 資料集上的實驗表明，相較於現有方法 FastGDBN 將最大收益提高了37.76倍並達到5,428倍的加速。;In modern IC design and manufacturing processes, VLSI testing plays a critical role in maintaining product reliability. In the context of conventional testing techniques struggling to achieve the desired defect coverage at a reasonable cost, researchers have applied data analysis techniques to develop testing methods that can further enhance defect coverage. One widely used method is known as Good-Die-in-Bad-Neighborhood (GDBN) method, which identifies high-risk dies by analyzing the distribution patterns of defective die clusters. This approach meets the stringent reliability requirements of fields such as automotive electronics and aerospace electronics. Recent studies have applied deep neural networks (DNNs) to achieve higher detection accuracy. However, the inherently high computational complexity of DNNs significantly reduces the efficiency of GDBN methods. To address the issue of detection efficiency degradation, we propose a highly parallelized neural network architecture called FastGDBN, which reduces inference time complexity to linear with respect to the number of wafers. Additionally, its end-to-end architecture minimizes CPU-GPU transfer overhead and eliminates redundant computations during the detection process of multiple dies from the same wafer. Experiments on the WM-811K dataset demonstrate that FastGDBN achieves a maximum gain increase of 37.76 times and an acceleration of 5,428 times compared to existing methods.
