AMD Tetrahedral Cages Cut Ray Tracing BVH Memory to 1.7GB
AMD has unveiled a research-driven approach for reducing the cost of ray tracing highly dynamic scenes by using tetrahedral cages to represent the deformation of dense animated geometry.
In a 1080p vegetation demonstration running on a Radeon RX 9070 XT, AMD reports that the technique can reduce ray-tracing acceleration-structure memory from as much as 80 GB to approximately 1.7 GB. Per-frame acceleration-structure update time also falls from more than 300 ms to approximately 3.3 ms.
The demonstration processes approximately 500 million dynamic triangles per frame for primary and shadow rays while maintaining more than 60 FPS at 1080p.
The underlying idea is to separate the animation of dense geometry from the expensive process of continuously rebuilding or updating its ray-tracing acceleration structures. Instead of deforming every triangle and updating its associated BVH, the approach applies animation to a much simpler tetrahedral cage surrounding the geometry.
🌿 Demonstration Handles 500 Million Dynamic Triangles #
AMD’s September 17 demonstration uses a large terrain environment containing approximately 25,000 independently animated plants.
At the highest level of detail, the vegetation represents roughly 2.8 billion triangles. Distance-based LOD selection reduces the active ray-traced workload to approximately 500 million dynamic triangles per frame.
The demonstration performs ray tracing for both primary and shadow rays and runs at 1080p on a Radeon RX 9070 XT. AMD reports performance exceeding 60 FPS.
The more significant result, however, is the reduction in acceleration-structure overhead.
| Metric | Conventional Dense Triangle BLAS | Tetrahedral Cage Approach |
|---|---|---|
| BVH memory | Up to 80 GB | ~1.7 GB |
| Per-frame BVH update | >300 ms | ~3.3 ms |
| Dynamic geometry | ~500 million triangles/frame | ~500 million triangles/frame |
| Output resolution | 1920×1080 | 1920×1080 |
These figures illustrate the fundamental difference between updating the entire animated geometry representation and updating a compact deformation structure.
🧊 How Tetrahedral Cages Work #
Ray-tracing pipelines rely on Bounding Volume Hierarchies (BVHs) to efficiently determine which pieces of geometry a ray can potentially intersect.
For independently animated meshes, conventional approaches generally need to account for changing vertex positions and update the relevant bottom-level acceleration structures as the geometry moves or deforms.
When thousands of objects are independently animated, the cost of maintaining those structures can become significant. The problem becomes particularly severe when each object contains dense geometry.
AMD’s tetrahedral cage approach changes the representation of the animation itself.
Separating deformation from dense geometry #
Instead of directly deforming the dense triangle mesh for acceleration-structure purposes, the technique places the geometry inside a relatively lightweight tetrahedral cage.
The dense geometry and its associated BVH can remain in their original rest-pose configuration. Animation is instead applied to the much simpler outer cage.
During ray intersection, the incoming ray is transformed into the coordinate space corresponding to the original rest pose. The existing static geometry and acceleration structure can then be reused while the cage deformation determines how the animated object should be interpreted.
This effectively moves much of the per-frame animation cost away from the triangle-level representation.
Why the approach scales differently #
With conventional dynamic geometry, the cost of maintaining acceleration structures is strongly influenced by the number of animated triangles.
With tetrahedral cages, the expensive acceleration structure can remain static while only the comparatively small deformation representation changes.
The result is a decoupling between animation complexity and total triangle count. A scene can therefore contain extremely dense geometry without requiring the entire BVH to be rebuilt or substantially updated every frame.
This is particularly useful for scenes containing many independently moving objects whose overall deformation can be represented adequately by a relatively simple cage.
🎯 Ideal Use Cases and Limitations #
Tetrahedral cages are not intended to replace conventional dynamic-geometry techniques in every rendering scenario.
Their greatest potential is in scenes containing large numbers of independently animated objects where the exact deformation of every vertex does not need to be represented with maximum precision.
Vegetation and environmental animation #
Large forests, grass fields, bushes, and other vegetation are natural candidates.
These scenes can contain enormous triangle counts while individual plants generally undergo relatively coherent deformations such as bending, swaying, or movement caused by environmental forces.
A tetrahedral cage can approximate these transformations without requiring the underlying dense geometry and its BVH to be updated at the same computational scale.
Crowds and distant characters #
Crowd rendering is another potential application. Large groups of characters can create substantial dynamic-geometry workloads when each character moves independently.
The approach can also be useful for distant characters where exact per-vertex animation is less important than maintaining visually plausible motion and efficient ray-tracing behavior.
Cases requiring precise deformation #
The technique becomes less suitable when animation requires highly detailed local deformation.
For example, if an object’s geometry undergoes complex changes that cannot be accurately represented by a relatively simple cage, the benefits can diminish. The same applies when mesh topology or triangle connectivity changes dynamically.
In those cases, traditional dynamic acceleration-structure updates may still be necessary.
The tetrahedral cage therefore functions as a specialized optimization rather than a universal replacement for existing dynamic ray-tracing pipelines.
📄 Research Behind the Technique #
The technology is based on the research paper Ray Tracing Massive Amounts of Animated Geometry, authored by Holger Gruen and colleagues.
The paper received the third-place Wolfgang Straßer Best Paper Award at the High-Performance Graphics 2026 conference.
The research focuses on the fundamental problem of efficiently ray tracing massive amounts of animated geometry. Rather than simply optimizing conventional BVH updates, the approach changes how dynamic deformation is represented so that the expensive acceleration structure can remain largely reusable.
That distinction is central to the technique’s potential scalability.
🛠️ AMD Is Working Toward Developer Integration #
AMD is also working on making the research more accessible to graphics developers.
The company is developing DirectX Raytracing (DXR) samples together with a header-only C++ library intended to help developers integrate tetrahedral-cage-based dynamic geometry into existing rendering pipelines.
A reusable implementation could make the technique easier to evaluate in production renderers and game engines without requiring developers to reproduce the research implementation from scratch.
However, the technology remains at the research and demonstration stage. It has not yet been established as a production technique in commercially released games.
🚀 A New Path for Dynamic Ray Tracing #
The most notable aspect of AMD’s tetrahedral cage research is the magnitude of the reported reduction in acceleration-structure overhead.
Reducing BVH memory from approximately 80 GB to 1.7 GB while cutting update time from more than 300 ms to roughly 3.3 ms fundamentally changes the resource requirements of scenes containing massive amounts of independently animated geometry.
The approach does not eliminate the computational cost of animation or ray tracing. Instead, it changes where that cost is paid by keeping dense geometry and its acceleration structure reusable while applying motion through a lightweight deformation cage.
If AMD’s approach proves practical across a broader range of dynamic scenes, tetrahedral cages could become another important technique for real-time ray tracing workloads involving dense vegetation, crowds, and other large-scale animated environments. For now, the technology remains a research-stage solution, but its reported results provide a compelling direction for reducing the memory and update costs associated with massive dynamic ray-traced scenes.