Pseudo-Labeling Based Semi-Supervised Learning for Signal Integrity Analysis of High-Bandwidth Memory (HBM) Interposer

In this article, a pseudolabeling (PL) based semisupervised learning method is proposed to identify the eye diagram distortion for accurately locating the signal integrity (SI) problems of high-bandwidth memory (HBM) silicon interposer channels. First, four main factors influencing the eye diagrams are presented, and 12 different eye diagram distortions are considered. The proposed convolutional neural network (CNN) and four different models are trained to identify these eye diagram distortions, and it is demonstrated that the proposed CNN exhibits good performance. Then, the PL method is applied to further improve the model performance. Finally, with the combination of the proposed CNN and PL method, the accuracy reaches up to 97.5% and becomes 32.3% higher than LeNet. Simultaneously, the graphic processing unit memory usage of the proposed model is 39.2% less than that of AlexNet. The proposed method provides an effective way for fast and accurately localizing the source of the SI problems for HBM interposer.