Optimizing meshoptimizer to process billions of triangles in minutes (2025)

Optimizing meshoptimizer to process the Zorah scene in minutes (2025)

NVIDIA's raytracing technology, RTX Mega Geometry, was released in early 2025, accompanied by a Zorah demo distributed as a ~100 GB Unreal Engine scene. This scene showcased clustered raytracing and Nanite's clustered LOD pipeline, allowing for the display of a highly detailed scene with full raytracing without using nanite proxy meshes.

In September 2025, NVIDIA released an open-source sample, vk_lod_clusters, featuring the Zorah scene as a glTF file. This sparked interest in improving support for hierarchical clustered LOD in meshoptimizer.

The Zorah scene contains 1.64 billion triangles, 18.9 billion triangles with instancing, and is 36.1 GB on disk. Importing this glTF file into Blender takes approximately 10 minutes and runs out of memory. Unreal Engine crashes in just under 5 minutes when importing the file.

The processing of the Zorah scene requires significant memory, and even 192 GB of RAM is insufficient for some users. However, running NVIDIA's sample code using 8 threads (--processingthreadpct 0.25) allowed successful processing, consuming approximately 180+ GB of RAM. Using 7 threads enabled partial use of the computer for about 30 minutes during the processing.

To build the hierarchical structure in the Zorah scene, meshoptimizer provides algorithms for clusterization, partitioning, and simplification. As of version 0.25, meshoptimizer contains two main clusterization algorithms.