The Gaussian Beam Tracing (GBT) method effectively addresses the singularity problems of traditional ray tracing in high-frequency acoustic simulations [1]. However, its application in large-scale complex environments is hindered by immense computational costs due to massive ray-surface interactions and beam summation operations. This study proposes a high-performance GPU-accelerated GBT solver that integrates CUDA Dynamic Parallelism with advanced spatial partitioning algorithms [2]. To overcome the performance bottleneck of ray-scene intersection tests, this work investigates three acceleration structures optimized for GPU architectures: Octrees, KD-trees, and Bounding Volume Hierarchies (BVH). Unlike traditional CPU-based Quadtree implementations, we evaluate these 3D structures based on construction overhead, memory access patterns, and thread divergence control. The results indicate that the BVH structure, when combined with dynamic parallelism, minimizes traversal steps and maximizes parallel thread occupancy. Furthermore, we utilize dynamic parallelism to handle the irregular workloads inherent in the Gaussian Beam Summation (GBS) process, allowing threads to dynamically spawn child kernels for computationally intensive regions. Simulation results in a complex city environment show that the proposed method achieves an overall pipeline acceleration of 112x compared to a single-threaded CPU baseline, with the GBS stage achieving a speedup of up to 817x. This approach enables efficient, high-precision environmental noise assessment.