Comparative Analysis of Virtualized Geometry versus Discrete Level of Detail in Unreal Engine 5

Sammanfattning

Real-time rendering performance has traditionally been managed using discrete Level of Detail (LOD) systems. While these systems aim to reduce the GPU load by simplifying geometry, they can introduce an additional burden on the CPU during camera visibility calculations or instance management. Unreal Engine 5 provides virtualized geometry through Nanite, a system designed to decouple geometric complexity from the CPU by shifting these visibility calculations to a GPU-driven data streaming pipeline. However, an uncertainty remains for developers regarding how performance scales across varying object densities, and whether a cross-over point exists where one pipeline becomes more efficient than the other on modern consumer hardware. This study investigated how the computational burden differs on the processor (measured in CPU Render Thread time) and GPU performance (measured as frame time and memory usage) between a Nanite-optimized setup and a traditional discrete LOD system in a high-density Unreal Engine 5 environment. To investigate this, a quantitative A/B experimental design was applied, utilizing a custom automated benchmarking tool to isolate geometry processing. Performance was observed in high-density scenes, ranging from 1,000 up to 256,000 object instances, using a heterogeneous set of 17 unique 3D models. Under the specific conditions tested, the empirical data did not indicate a performance cross-over point within the tested range, as Nanite maintained lower or comparable computation times across all levels of object density. Instead, a divergence point was discovered between 4,000 and 16,000 objects. Beyond this threshold, although the traditional LOD system remained interactive on the tested hardware, Nanite proved to be significantly more efficient at scaling. At the maximum observed density of 256,000 objects, Nanite recorded an average CPU Render Thread time of 6.61 ms, compared to 12.24 ms for the traditional system. GPU frame times at this density were observed at 4.88 ms for Nanite and 11.04 ms for traditional LODs. Furthermore, Nanite exhibited superior frame-pacing stability (measured as 99th percentile frame times), effectively reducing micro-stuttering with a recorded 10.11 ms compared to 15.71 ms for the traditional system at the highest density. These observations suggest that developers can utilize virtualized geometry for static 3D models on modern hardware without incurring performance penalties at lower densities. However, the system lockups observed at 1,024,000 objects indicate that managing spatial data (transform arrays) for millions of objects on the CPU remains a significant scalability limitation in both architectures.