AMD Tetrahedral Cages Cuts Ray Tracing VRAM by 47x
AMD has introduced a real-time ray tracing technique called tetrahedral cages through its GPUOpen developer platform, targeting one of the major challenges in highly dynamic, high-density scenes: the memory and update costs of ray tracing acceleration structures.
In AMD’s demonstration, a scene containing approximately 25,000 independently animated plants uses roughly 1.7GB of BVH VRAM, compared with up to 80GB for a conventional dense-triangle BLAS approach. Per-frame BVH update time also falls from more than 300 ms to approximately 3.3 ms.
That represents an approximately 47x reduction in BVH memory consumption and roughly a 90x reduction in acceleration-structure update time.
The technology is still a research and developer demonstration rather than a production-ready replacement for conventional animation pipelines. Its primary target is dense geometry that does not require detailed per-vertex deformation, such as grass, foliage, and similar environmental assets.
📊 AMD Demonstrates Major BVH Memory Reduction #
AMD’s technical demonstration uses a large terrain scene containing approximately 25,000 independently animated plants.
Each plant can respond to locally different wind-field simulations, creating a highly dynamic environment with substantial geometric complexity.
At maximum level of detail, the scene contains approximately 2.8 billion triangles. After level-of-detail selection, around 500 million animated plant triangles actively participate in ray tracing during each frame.
The demonstration runs at 1080p with primary and shadow rays and achieves more than 60 FPS on a Radeon RX 9070 XT.
The key result, however, is not simply the number of triangles rendered. AMD’s technique is designed to reduce the cost of maintaining the bounding volume hierarchy used by ray tracing.
For the same scene, AMD reports the following comparison:
| Ray tracing approach | BVH VRAM usage | BVH update time |
|---|---|---|
| Conventional dense-triangle BLAS | Up to 80GB | More than 300ms/frame |
| Tetrahedral Cages | Approximately 1.7GB | Approximately 3.3ms/frame |
The resulting difference is substantial:
- VRAM usage: approximately 47x lower
- BVH update time: approximately 90x faster
These reductions target two major bottlenecks that become increasingly difficult to manage as animated geometry becomes denser.
🧩 How Tetrahedral Cages Work #
Traditional real-time ray tracing requires the acceleration structures associated with animated geometry to be updated as the underlying meshes deform.
For highly detailed models, this can become expensive because the cost of updating the acceleration structure is closely related to the amount of geometry being animated.
Foliage provides an extreme example. A forest may contain millions or billions of triangles while many individual plants continuously deform in response to wind.
Updating every animated triangle and rebuilding or refitting the associated acceleration structures each frame can consume substantial compute resources and VRAM.
Tetrahedral cages approach the problem differently by separating the complexity of the rendered geometry from the complexity of its animation representation.
Separating Dense Geometry From Animation #
The basic concept is to enclose a high-density animated mesh within a significantly simpler deformable tetrahedral cage.
The dense internal geometry and its associated BVH can remain static and reusable. Instead of updating every vertex of the original mesh, the system updates the lower-complexity tetrahedral cage.
Conceptually, the rendering process works as follows:
- The high-density model is enclosed by a simplified tetrahedral cage.
- The detailed geometry and its acceleration structure remain static.
- The cage is deformed to represent the object’s animation.
- The ray intersects the animated cage.
- The ray is transformed into the corresponding frame of the static internal geometry.
- The existing static acceleration structure is then used to perform the detailed ray query.
This effectively moves the per-frame update workload away from the original triangle count and toward the considerably smaller cage representation.
Why This Matters for Dense Geometry #
With conventional approaches, a highly detailed mesh can require expensive acceleration-structure updates even when its animation is relatively simple.
Tetrahedral cages exploit the fact that some objects have enormous geometric complexity but relatively low-frequency deformation.
For example, a plant may contain thousands of triangles but its overall movement under wind can be represented reasonably well by deforming a much simpler cage.
The technique therefore attempts to make animation-update complexity more closely track the complexity of the deformation rather than the complexity of the rendered mesh.
🌿 Best Suited to Foliage and Similar Assets #
Tetrahedral cages are not intended to replace every existing animation technique.
The primary limitation comes from the same simplification that produces the performance benefits: animation is represented through a relatively coarse cage.
That means developers lose some of the fine-grained control available with direct per-vertex animation.
This makes the technique less suitable for assets where small geometric deformations are visually important.
Character animation, facial deformation, cloth simulation, and other assets that require precise vertex-level control may not be ideal targets.
Instead, the technology is better suited to highly detailed objects whose deformation can tolerate some approximation.
Potential examples include:
- Grass
- Trees and foliage
- Dense vegetation
- Environmental assets
- Other high-triangle-count objects with relatively simple deformation
In these scenarios, the geometry may be extremely expensive to represent in a ray tracing acceleration structure, while the actual animation can be described using a comparatively small number of cage elements.
DirectX Raytracing Compatibility #
AMD has also indicated that tetrahedral cages can be combined with Microsoft DirectX Raytracing (DXR).
This makes the technique relevant to existing real-time ray tracing pipelines rather than requiring an entirely separate rendering architecture.
However, integration into a production engine would still require appropriate tooling, asset-processing workflows, and runtime support.
⚙️ The Trade-Off Between Accuracy and Efficiency #
The approximately 47x reduction in VRAM usage and 90x improvement in update time come from changing how animated geometry is represented.
The approach therefore involves an explicit trade-off.
A conventional pipeline provides direct control over detailed vertex deformation but can become increasingly expensive as geometric complexity increases.
Tetrahedral cages reduce the amount of animation data that must be updated, but the resulting deformation is constrained by the cage representation.
The technique is consequently best viewed as a specialized optimization rather than a universal replacement for conventional skinned or vertex-based animation.
For suitable assets, the reduction in acceleration-structure overhead could be substantial. For assets requiring precise local deformation, the approximation may introduce unacceptable visual differences.
The appropriate choice therefore depends on the relationship between geometric complexity and animation complexity in a particular scene.
🧪 Current Status: Research and Developer Demonstration #
Despite the impressive numbers in AMD’s demonstration, tetrahedral cages remain in the research and developer-demo stage.
AMD has not announced a schedule for integrating the technology into commercial games or shipping it as a standard driver-level feature. There is also no confirmed timeline for broader adoption through standardized graphics development toolchains.
Moving from a research demonstration to production use would require more than demonstrating favorable benchmark numbers.
Engine developers would need practical tools for generating and authoring cages, asset pipelines would need to support the representation, and rendering engines would need appropriate runtime integration.
Visual quality would also need to be evaluated across a much wider range of real-world assets and animation patterns.
🚧 Adoption Will Depend on Workflow Integration #
The most important question is therefore not whether tetrahedral cages can reduce BVH overhead in the demonstrated scenario. AMD’s data shows that they can achieve dramatic reductions under the tested conditions.
The larger question is how easily the technique can be integrated into commercial rendering workflows.
Potential adoption factors include:
- Cage-generation and authoring complexity
- Compatibility with existing animation systems
- Runtime integration requirements
- Visual differences caused by coarse deformation
- Asset-processing overhead
- Support across rendering APIs and engines
- Performance across different GPU architectures and scene types
If these practical challenges can be addressed, the technique could provide a useful way to make highly detailed dynamic environments more compatible with real-time ray tracing.
📝 Final Takeaways #
AMD’s tetrahedral cages technology attacks a specific but increasingly important problem in real-time ray tracing: the cost of updating acceleration structures for extremely dense animated geometry.
In AMD’s approximately 25,000-plant demonstration, the technique reduces reported BVH VRAM usage from up to 80GB to approximately 1.7GB and cuts update time from more than 300 ms to approximately 3.3 ms per frame.
The underlying idea is to keep the detailed geometry and its BVH static while using a much simpler tetrahedral cage to represent animation. This decouples the per-frame update workload from the triangle count of the original model.
The approach is particularly promising for dense environmental geometry such as foliage, where millions or billions of triangles can undergo relatively simple deformation.
However, the technique remains a research demonstration. Its coarse deformation model makes it less suitable for applications requiring precise per-vertex animation, and widespread adoption will depend on tooling, engine integration, visual quality, and real-world performance beyond AMD’s demonstration scene.
For now, tetrahedral cages represent an interesting direction for reducing the memory and update overhead of ray tracing in highly complex dynamic environments rather than a general-purpose replacement for existing animation techniques.