Characterization and Analysis of the 3D Gaussian Splatting Rendering Pipeline

Novel view synthesis, a task generating a 2D image frame from a specific viewpoint within a 3D object or scene, plays a crucial role in 3D rendering. Neural Radiance Field (NeRF) emerged as a prominent method for implementing novel view synthesis, but 3D Gaussian Splatting (3DGS) recently began to emerge as a viable alternative. Despite the tremendous interest from both academia and industry, there has been a lack of research to identify the computational bottlenecks of 3DGS, which is critical for its deployment in real-world products. In this work, we present a comprehensive end-to-end characterization of the 3DGS rendering pipeline, identifying the alpha blending stage within the tile-based rasterizer as causing a significant performance bottleneck. Based on our findings, we discuss several future research directions aiming to inspire continued exploration within this burgeoning application domain.