Bao Rong Chang

This review examines the transformative interplay between computing chips and Artificial Intelligence (AI), driving a revolution across various industries. First, the broader artificial intelligence and semiconductor ecosystem is analyzed, including hardware manufacturers, software frameworks, and system integration. Next, the development prospects are examined, revealing current challenges such as power consumption, manufacturing complexity, supply chain constraints, and ethical considerations. Further discussion focuses on cloud-edge collaboration in relation to system architecture and workload allocation strategies. Then, cutting-edge AI technologies are analyzed, and key insights are summarized. Finally, the overall trends in artificial intelligence and the chip industry are summarized, clearly presenting the findings for the future and making a unique contribution to this review.

The die bonding process significantly impacts the yield and quality of IC packaging, and its quality detection is also a critical image sensing technology. With the advancement of machine automation and increased operating speeds, the misclassification rate in die bond image inspection has also risen. Therefore, this study develops a high-speed intelligent vision inspection model that slightly improves classification accuracy and adapts to the operation of new-generation machines. Furthermore, by identifying the causes of die bonding defects, key process parameters can be adjusted in real time during production, thereby improving the yield of the die bonding process and substantially reducing manufacturing cost losses. Previously, we proposed a lightweight model named DSGβSI-YOLOv7-tiny, which integrates depthwise separable convolution, Ghost convolution, and a Sigmoid activation function with a learnable β parameter. This model enables real-time and efficient detection and prediction of die bond quality through image sensing. We further enhanced the previous model by incorporating an SE layer, ECA-Net, Coordinate Attention, and a Small Object Enhancer to accommodate the faster operation of new machines. This improvement resulted in a more lightweight architecture named DSGβSI-SECS-YOLOv7-tiny. Compared with the previous model, the proposed model achieves an increased inference speed of 294.1 FPS and a Precision of 99.1%.

In the previous work, we proposed LWGSE-YOLOv4-tiny and LWDSG-ResNet18, leveraging depthwise separable and Ghost Convolutions for fast self-driving control while achieving a detection speed of 24.9 FPS. However, the system fell short of Level 4 autonomous driving safety requirements. That is, the control response speed of object detection integrated with steering angle prediction must exceed 39.2 FPS. This study enhances YOLOv11n with dual convolution and RepGhost bottleneck, forming DuCRG-YOLOv11n, significantly improving the object detection speed while maintaining accuracy. Similarly, DuC-ResNet18 improves steering angle prediction speed and accuracy. Our approach achieves 50.7 FPS, meeting Level 4 safety standards. Compared to previous work, DuCRG-YOLOv11n boosts feature extraction speed by 912.97%, while DuC-ResNet18 enhances prediction speed by 45.37% and accuracy by 12.26%.

The die bonding process is one of the most critical steps in the front-end semiconductor packaging process, as it significantly affects the yield of the entire IC packaging process. This research aims to find an efficient, intelligent vision detection model to identify whether each chip correctly adheres to the IC substrate; by utilizing the detection model to classify the type of defects occurring in the die bond images, the engineers can analyze the leading causes, enabling timely adjustments to key machine parameters in real-time, improving the yield of the die bond process, and significantly reducing manufacturing cost losses. This study proposes the lightweight Yolov7-tiny model using Depthwise-Separable and Ghost Convolutions and Sigmoid Linear Unit with β parameter (DSGβSI-Yolov7-tiny), which we can apply for real-time and efficient detection and prediction of die bond quality. The model achieves a maximum FPS of 192.3, a precision of 99.1%, and an F1-score of 0.97. Therefore, the performance of the proposed DSGβSI-Yolov7-tiny model outperforms other methods.

Molding injects a molding compound into a mold to form a protective shell around the wafer. During the injection process, overflow may occur, leading to mold flash, which reduces yield and causes significant manufacturing cost losses. This paper proposes a deep-learning-based method for detecting and predicting the occurrence of mold flash probability to address this issue. First, the paper conducts random forest importance analysis and correlation analysis to identify the key parameters that significantly impact mold flash. This paper uses these key parameters as input signals for the prediction model. The paper introduces an HLGA Transformer to construct an ensemble meta-learning model that predicts the probability of molding defects, achieving a prediction accuracy of 98.16%. The ensemble meta-learning approach proposed in this paper outperforms other methods in terms of performance. The model predictions can be communicated to the system in real time, allowing it to promptly adjust critical machine operation parameters, thereby significantly improving the molding process yield and reducing substantial manufacturing cost losses.

In wafer testing, the test probe needs to contact the surface of the semiconductor wafer to measure electrical parameters such as resistance, capacitance, and current. The probe mark damage frequently occurs on the machine during wafer testing. This phenomenon causes inaccuracies in electrical parameter testing, significantly reducing the yield in IC packaging processes and resulting in substantial manufacturing cost losses. Therefore, this study proposed an effective probe mark damage detection and prediction method to prevent significant yield reduction due to inaccurate testing. This study uses importance analysis from random forests and correlation analysis to identify the critical parameters influencing probe mark damage. Introducing these parameters into a Sparse Transformer with hybrid normalization can successfully train an intelligent model for predicting the occurrence of probe mark damage. The model accurately predicts the probability of probe mark damage and promptly adjusts machine parameters to avoid inaccuracies in electrical parameter settings. The proposed approach can outperform other methods, achieving two very high accuracies of 95.1% (at room temperature) and 93.5% (at high temperature) and significantly reducing the occurrence of large-area probe mark damage.

As the user’s behavior changes at any time with cloud computing and network services, abnormal server resource utilization traffic will lead to severe service crashes and system downtime. The traditional single anomaly detection model cannot handle the rapid failure prediction ahead. Therefore, this study proposed ensemble learning combined with model-agnostic meta-reinforcement learning called ensemble meta-reinforcement learning (EMRL) to implement self-adaptive server anomaly detection rapidly and precisely, according to the time series of server resource utilization. The proposed ensemble approach combines hidden Markov model (HMM), variational autoencoder (VAE), temporal convolutional autoencoder (TCN-AE), and bidirectional long short-term memory (BLSTM). The EMRL algorithm trains this combination with several tasks to learn the implicit representation of various anomalous traffic, where each task executes trust region policy optimization (TRPO) to quickly adapt the time-varying data distribution and make rapid decisions precisely for an agent response. As a result, our proposed approach can improve the precision of anomaly prediction by 2.4 times and reduce the model deployment speed by 5.8 times on average because a meta-learner can immediately be applied to new tasks.

Our previous study utilized importance analysis, random forest, and Barnes–Hut t-SNE dimensionality reduction to analyze critical dicing parameters and used bidirectional long short-term memory (BLSTM) to predict wafer chipping occurrence successfully in a single dicing machine. However, each dicing machine of the same type may produce unevenly distributed non-IID dicing signals, which may lead to the undesirable result that a pre-trained model trained by dicing machine #1 could not effectively predict chipping occurrence in dicing machine #2. Therefore, regarding the model robustness, this study introduces an ensemble meta-learning-based model that can evaluate many dicing machines for chipping prediction with high stability and accuracy. This approach constructs several base learners, such as the hidden Markov model (HMM), the variational autoencoder (VAE), and BLSTM, to form an ensemble learning. We use model-agnostic meta-learning (MAML) to train and test the ensemble learning model by several prediction tasks from machine #1. After MAML learning, we call the trained model a meta learner. Then, we successfully apply a retrieved data set from machine #2 to the meta learner for training and testing wafer chipping occurrence in this machine. As a result, our contribution to the robust chipping prediction on cross-machines can improve the yield of wafer dicing with a prediction accuracy of 93.21%, preserve the practical wearing of dicing kerfs, and significantly cut wafer manufacturing costs.

Chip contour detection aims to detect damaged chips in chip slots during IC packaging and testing using vision facilities. However, the operation speed of the new chip transportation machine is too fast, and the current chip contour detection models, such as Yolov5, M3-Yolov5, FGHSE-Yolov5, and GSEH-Yolov5, running on the embedded platform, Jetson Nano, cannot detect chip contours in a timely manner. Therefore, there must be a rapid response for chip contour detection. This paper introduces the DSGSE-Yolov5s algorithm, which can accelerate object detection and image recognition to resolve this problem. Additionally, this study makes a performance comparison between the different models. Compared with the traditional model Yolov5, the proposed DSGSE-Yolov5s algorithm can significantly promote the speed of object detection by 132.17% and slightly increase the precision by 0.85%. As a result, the proposed approach can outperform the other methods.