This paper proposes a novel scheduling and analysis method for Directed Acyclic Graph (DAG)-structured GPU tasks to enhance real-time predictability. It introduces techniques like parallelism scaling and node segmentation to reduce task execution time and provide a safe makespan bound without relying on kernel priorities.