This blog post introduces NVIDIA OptiX 9's new RTX Mega Geometry features, specifically Cluster Acceleration Structures (CLAS) and Cluster Templates, designed to significantly enhance real-time ray tracing performance for dynamic, high-density scenes. Traditional methods struggle with the frequent rebuilding of geometry acceleration structures required for animated content, especially subdivision surfaces. CLAS addresses this by enabling efficient reuse of small triangle clusters, either across frames or spatially within the same frame. This approach leads to acceleration structure rebuilds that are 10x-100x faster than full rebuilds, making real-time adaptive tessellation of complex surfaces feasible. The post also details how this pipeline integrates with DLSS Ray Reconstruction for stable, high-quality visual output, providing a robust solution for rendering realistic animated geometry in real-time.
Summarizing the Text on Fast Ray Tracing of Dynamic Scenes:
The Challenge of Dynamic Scenes: Real-time ray tracing of highly detailed, dynamic geometry (where geometry changes every frame) is difficult because the geometry acceleration structures (search trees) must be constantly updated, often taking longer than the tracing itself.
Limitations of Existing Methods: While a "refit" operation exists for scenes where only vertex positions change, it's slower if vertices move too much and doesn't allow for changes in triangle count or connectivity.
Introducing Cluster Acceleration Structures (CLAS): NVIDIA OptiX 9 introduces CLAS, a new data structure where users define small clusters of triangles as inputs. These CLASes then become inputs for a higher-level Cluster Geometry Acceleration Structure (Cluster GAS).
Benefits of CLAS: CLAS offers the speed benefits of a refit while retaining most of the flexibility and trace performance of a full rebuild. It reuses patterns at the cluster level, performing a full rebuild only over these clusters.
Types of Cluster Reuse:
Temporal Reuse: The same clusters can be reused across frames, even with animated vertices (similar to GAS refit).
Spatial Reuse: The same cluster pattern (topology) can be reused spatially within the same frame, even if positions differ. This is particularly effective for dynamic tessellation of subdivision surfaces.
Subdivision Surfaces: Catmull-Clark subdivision surfaces (subds), common in film, use a lightweight control mesh and mathematical rules to create smooth limit surfaces with unlimited microtriangle density, allowing for adaptive tessellation based on camera view.
Adaptive Tessellation Pipeline: A proposed real-time ray tracing pipeline using OptiX with cluster geometry involves:
Animating the subd control mesh.
Tessellating the limit surface into clusters (around one triangle per pixel) and saving their descriptions/vertices.
Instantiating CLASes for each cluster using a single, device-memory-driven function call.
Building higher-level GAS and IAS, which is 10x-100x faster than building directly over microtriangles.
Launching OptiX to shoot rays and shade, with OptiX providing cluster ID and hit triangle vertices.
Cluster Templates: These are reusable cluster definitions in the OptiX API, storing everything about a CLAS except its vertices. They can be quickly instantiated with new vertices of the same topology. OptiX also provides optimized API calls for grid-based templates.
Overall Speedup: Cluster Templates can speed up acceleration structure rebuilds by an order of magnitude compared to placing all triangles into a single flat GAS.
DLSS Ray Reconstruction (DLSS-RR) Integration: To achieve visually smooth images at high frame rates (e.g., one sample per pixel), DLSS-RR can replace traditional accumulation loops. It uses motion vectors and other input layers for internal accumulation, producing stable images, though not converging to a reference image.
Key Takeaway: OptiX 9's CLAS and RTX Mega Geometry fundamentally improve the efficiency of ray tracing dynamic, high-density geometry, especially subdivision surfaces, by optimizing acceleration structure updates through intelligent cluster reuse. This, combined with DLSS-RR, enables high-quality real-time rendering of complex animated scenes.