Chip surface defect detection is a critical research task and a vital component of integrated circuit quality inspection. With advancements in artificial intelligence, deep learning-based defect detection methods have become prevalent. However, due to the complex morphology of chip defects and their susceptibility to environmental background interference, precisely detecting micro-scale and multi-scale chip defects in large-scale, high-resolution images remains a significant challenge. In this paper, we propose a Multi-Scale Dense Network (MSDNet) for chip surface defect segmentation. The proposed MSDNet utilizes an encoder-decoder framework, incorporating multi-scale convolution modules, attention modules, and a dense node module to enhance defect segmentation performance. Additionally, we construct a chip defect dataset and conduct extensive experimental verifications. The experimental results demonstrate that the proposed MSDNet achieves an 85.01% defect segmentation accuracy on the chip defect dataset. Compared to the baseline, our proposed MSDNet significantly improves defect segmentation performance and ensures more precise segmentation of defect details.