AMD tetrahedral cage technique cuts ray tracing BVH memory from 80 GB to 1.7 GB

September 20, 2026 at 07:50 PM2 min

This digest was compiled by AI from multiple sources — links to the originals are below.

AMD demonstrated a tetrahedral cage technique that reduces ray tracing BVH memory usage from 80 GB to 1.7 GB on a Radeon RX 9070 XT. The method ray-traces 500 million animated triangles at over 60 FPS at 1080p. BVH update time drops from over 300 ms to 3.3 ms per frame.

Key Facts

AMD's tetrahedral cage technique reduces BVH memory usage from up to 80 GB to approximately 1.7 GB in a foliage demonstration.
BVH update time per frame drops from over 300 ms to about 3.3 ms using the technique.
The demonstration ray-traces around 500 million animated triangles at over 60 FPS at 1080p on a Radeon RX 9070 XT.
The technique was published in a GPUOpen article on September 17 by AMD Fellow Holger Gruen.
The method encloses dense geometry in deformable tetrahedral cages, keeping the original geometry and BVH static and reusable.

The Technique

AMD's approach encloses dense geometry within comparatively simple deformable tetrahedral cages. Instead of updating every individual vertex and its associated acceleration structure, the technique animates the surrounding cage while keeping the original dense geometry and its BVH static and reusable. Rays are transformed back into the geometry's original reference frame before intersection testing.

Performance Results

In a large-scale terrain scene with approximately 25,000 independently animated plants, the technique ray-traces around 500 million animated triangles per frame after LOD selection. The scene runs at over 60 FPS at 1080p on a Radeon RX 9070 XT. BVH memory usage drops from up to 80 GB to approximately 1.7 GB, and BVH update time per frame falls from over 300 ms to about 3.3 ms. These figures represent roughly 47 times less acceleration-structure memory and a more than 90-fold reduction in BVH update time.

Trade-offs

Tetrahedral cages sacrifice fine-grained control over individual vertex animations. The technique is particularly suitable for dense foliage, grass, crowds, and distant characters rather than animations requiring precise deformation or changes in mesh topology.