As a cornerstone of the semiconductor industry, the surface quality of wafers directly impacts chip yield and performance. Minute defects, such as chipping, edge chipping, and positive chipping, can lead to circuit failure. Traditional optical inspection methods, which rely on human operators, suffer from low efficiency, subjectivity, and operator fatigue, making them inadequate for the high-precision and high-efficiency demands of modern semiconductor manufacturing. This paper introduces a real-time detection method based on the YOLOv8 object detection algorithm and the NVIDIA TensorRT inference accelerator. This method leverages YOLOv8's superior detection accuracy and speed, combined with TensorRT's model optimization and acceleration capabilities, achieving a mean average precision mAP of up to 98.2% and an inference speed of 217 FPS. Experimental results demonstrate that this system can efficiently and accurately locate and classify wafer defects on real-time production lines, providing a reliable embedded solution for improving quality control processes in semiconductor manufacturing.

defect detection; wafer surface; YOLOv8; TensorRT;