The study of decision-making in large-scale game environments is a crucial domain within artificial intelligence, possessing substantial implications for practical applications. Nevertheless, the lack of comprehensive, realistic game environments and associated datasets has limited progress in this field. To address this and promote research on this important problem, we introduce the Large-Scale Auction (LSA) Benchmark derived from online advertising, a rapidly expanding industry worth $626.8 billion in 2023. The LSA Benchmark consists of an environment and the corresponding dataset. The LSA Environment is augmented with the deep generative model to reduce the gap between the simulation environment and reality while avoiding the risks of sensitive data exposure. The LSA Dataset comprises over 500 million records, totaling 40 GB in size, which contains trajectories with 50 diverse agents competing with each other, for effective offline training. We evaluate different types of existing algorithms in the LSA Environment. We hope the LSA benchmark can promote the development of decision-making in large-scale games.