AMD Demo Cuts Ray-Tracing BVH Memory From 80 GB to 1.7 GB

AMD has published new measurements from its tetrahedral-cage ray-tracing research, showing a large reduction in acceleration-structure memory and update cost for a deliberately extreme animated-vegetation scene. The September 17 GPUOpen update demonstrates roughly 25,000 independently animated plants on a Radeon RX 9070 XT at 1080p.

At the highest level of detail the plants represent about 2.8 billion triangles. After level-of-detail selection, AMD says roughly 500 million animated plant triangles are ray traced per frame with primary and shadow rays at more than 60 FPS. The important result is not simply the triangle count: AMD reports that its tetrahedral-cage representation needs about 1.7 GB of BVH memory and roughly 3.3 ms for BVH updates per frame in this demo.

Conventional ray tracing becomes awkward when large amounts of geometry deform independently. Once vertex positions change, the acceleration structure describing that geometry may need to be updated as well. With many dense animated objects, both the stored BVHs and the work needed to keep them current can become expensive.

AMD’s method decouples the animation from the dense mesh. During preprocessing, a lower-resolution tetrahedral cage encloses the original geometry and the mesh is divided into small pieces associated with those tetrahedra. Static mini-BLAS structures for the detailed rest-pose geometry can then be reused. At runtime, the cage deforms while rays entering an animated tetrahedron are transformed back into the corresponding rest-pose space for intersection with the static geometry.

That changes what scales with animation complexity. Instead of updating acceleration structures for every dense animated triangle, the per-frame deformation cost depends primarily on the much smaller cage. Multiple unique deformations can still share the detailed rest-pose geometry and its static acceleration data.

A conventional dense approach requiring up to 80 GB of BVH memory is plainly outside the local memory capacity of a Radeon RX 9070 XT. AMD’s roughly 1.7 GB cage result brings the acceleration data for this particular stress test into a range that can coexist with the rest of a real-time rendering workload, although the complete application still needs memory for geometry, textures, render targets and other resources.

The update-time reduction tackles a different bottleneck. More than 300 ms of acceleration-structure work cannot fit inside a real-time frame budget, whereas AMD’s reported 3.3 ms is at least compatible with a 60 FPS-class frame budget if the rest of the renderer also fits. The published 60+ FPS demo is therefore stronger evidence than quoting the BVH update figure alone, but it still describes AMD’s controlled research workload rather than a shipping game benchmark.

The cage is an approximation of the original animation. AMD describes the deformation as piecewise linear, with a direct trade-off between cage resolution, memory and animation fidelity. A coarse cage is cheaper but follows fine motion less accurately; a denser cage can better approximate the intended deformation while giving back some of the efficiency.

That makes the method especially interesting for connectivity-preserving motion such as swaying vegetation, grass, crowds, distant characters and animation level-of-detail systems. It is less naturally suited to topology changes or cases where very small-scale deformation is visually critical. This is an optimization technique with a content domain, not a replacement for every animated BLAS workload.

AMD says it is working on DirectX Raytracing samples and a header-only C++ library for building tetrahedral cages for skinned, keyframe-animated and static objects. The research can also coexist with newer acceleration-structure techniques rather than requiring an entirely separate ray-tracing pipeline.

For PC games, the consequential question is whether developers can integrate the method without unacceptable authoring, preprocessing or visual-quality costs. If that work proves practical, dense reactive foliage and other uniquely animated geometry could become substantially less hostile to ray-traced rendering. Until production integrations exist, the current numbers should remain what they are: unusually strong results from a focused AMD research demonstration.