Entity alignment (EA) in knowledge graphs (KGs) aims to identify equivalent entities across different graphs, serving as a critical step in cross-lingual KG integration. However, existing EA methods based on graph neural networks (GNNs) suffer from scalability issues and lack interpretability. To address these challenges, this paper proposes a lightweight, unsupervised, and efficient entity alignment framework, LWUE. The encoder in LWUE retains only graph convolution operations for feature embedding, significantly reducing spatiotemporal complexity. The decoder maximizes homogeneity by minimizing the Dirichlet energy of entity embeddings and employs the Hungarian or Sinkhorn algorithm for alignment. Experimental results demonstrate that LWUE achieves competitive performance on public datasets, completing alignment tasks in a short time and exhibiting significant advantages in efficiency.