AMD Ray Tracing Technique Cuts BVH Memory From 80GB to 1.7GB

AMD has demonstrated a new ray tracing technique that can sharply reduce the memory and processing cost of handling extremely complex animated scenes.

The research uses what AMD calls tetrahedral cages to simplify how animated geometry interacts with ray tracing acceleration structures. In a large demonstration scene, the approach reduced BVH memory usage from as much as 80 GB to around 1.7 GB while cutting update time from more than 300ms per frame to roughly 3.3ms.

The same demo rendered around 500 million animated triangles per frame at more than 60 FPS and 1080p on a Radeon RX 9070 XT.

The technique keeps dense geometry static

Modern ray tracing relies heavily on Bounding Volume Hierarchies, commonly called BVHs.

These acceleration structures help the GPU determine which parts of a scene a ray might intersect without testing every triangle individually.

The problem becomes much harder when large amounts of geometry are animated independently.

In a traditional approach, each animated object may need its geometry and BVH updated as it moves or deforms. When a scene contains thousands of highly detailed animated objects, those updates can consume enormous amounts of memory and processing time.

AMD's approach works differently.

Instead of rebuilding acceleration structures for every detailed object, the dense geometry is enclosed inside a simpler deformable tetrahedral cage.

The cage is animated while the original high detail geometry and its BVH remain static and reusable.

When a ray enters the animated cage, it is transformed back into the original reference space before intersection testing takes place.

Metric	Conventional approach	Tetrahedral cage approach
BVH memory usage	Up to 80 GB	About 1.7 GB
BVH update time	More than 300ms per frame	About 3.3ms
Animated triangles after LOD	Around 500 million	Around 500 million
Test GPU	Radeon RX 9070 XT	Radeon RX 9070 XT
Performance target	Not practical at this scale	More than 60 FPS at 1080p

The difference represents roughly 47 times lower acceleration structure memory usage and more than a 90 times reduction in BVH update time.

AMD tested the method with 25,000 animated plants

The demonstration scene contains about 25,000 independently animated plants.

At their maximum level of detail, the vegetation represents roughly 2.8 billion triangles.

After level of detail selection reduces the amount of visible geometry that needs to be processed, around 500 million triangles remain active per frame.

Those triangles are then ray traced using primary and shadow rays while the scene runs at more than 60 FPS at 1080p on the RX 9070 XT.

That workload shows why reducing BVH update costs matters.

Without a more efficient technique, maintaining acceleration structures for that much independently moving geometry would consume an impractical amount of memory and processing time.

The biggest gains are likely for foliage and crowds

Tetrahedral cages are not ideal for every kind of animation.

The technique sacrifices some fine control over individual vertices because the dense geometry is deformed indirectly through the surrounding cage.

That makes it better suited to objects where approximate deformation is acceptable.

Dense grass, trees, foliage, large crowds and distant characters are strong candidates.

Animations that require highly precise mesh deformation or major topology changes may still need conventional techniques.

This means the technology is more likely to complement existing ray tracing systems rather than replace all forms of animated geometry processing.

The research is not yet part of shipping games

AMD's work is still a research project rather than a feature currently available in commercial games.

The underlying paper on ray tracing massive amounts of animated geometry received third place in the Wolfgang Straßer Best Paper Award at High Performance Graphics 2026.

AMD is also developing DirectX Raytracing samples and a header only C++ library that could make the technique easier for developers to experiment with.

If the method eventually reaches production engines, it could allow developers to build much denser animated environments without requiring enormous amounts of VRAM.

The most important result is not simply the 500 million triangle figure. Reducing BVH memory from 80 GB to about 1.7 GB addresses one of the practical limitations that makes large animated ray traced scenes difficult to render efficiently today.