Virtualized Geometry Systems: Architecture, Theory, and Implementation
A Principal Engineer's Guide to Production-Grade Micro-Polygon Rendering
"The geometry problem is not one of bandwidth alone, it is a problem of indirection, coherence, and the fundamental mismatch between the triangle and the pixel."
Abstract
Modern game engines face an intractable triangle throughput ceiling: GPU rasterizers are designed around polygons covering several pixels, yet cinematic assets routinely produce triangles smaller than a single pixel at runtime distances. Classical Level-of-Detail (LOD) pipelines fail at this scale due to popping artifacts, authoring overhead, and the inability to leverage the full spatial resolution of sculpted meshes. This paper covers Virtualized Geometry Systems (VGS), the family of techniques that includes Epic Games' Nanite (Unreal Engine 5), Ubisoft's micropolygon pipeline, and related academic precursors such as Reyes, Progressive Meshes, and Streaming Mesh systems. Topics include cluster-based hierarchical level-of-detail, screen-space error metrics, BVH/DAG visibility culling, software rasterization fallback paths, virtual shadow maps, and GPU-driven draw dispatch. Included is a concrete implementation blueprint targeting a C++23/Vulkan 1.3 engine, with pseudocode, data layout specifications, and complexity analyses.
Keywords: virtualized geometry, LOD, BVH, cluster hierarchy, software rasterization, meshlet, screen-space error, GPU-driven rendering, visibility buffer, nanite
