Real-time ray tracing has a known cost problem with animated scenes, and AMD just published a fix that cuts it by nearly two orders of magnitude. A bounding volume hierarchy (BVH) is the data structure a ray tracer uses to figure out which surfaces a ray might hit — cheap to build once for static geometry, expensive to rebuild every frame when thousands of objects are independently deforming, which is exactly what a wind-blown field of plants does.
AMD Fellow Holger Gruen’s fix, published on GPUOpen on September 17 and covered in an ACM SIGGRAPH Blog roundup by Jon Peddie on October 29, encloses dense geometry inside a simple deformable structure of tetrahedra — a “cage.” The cage moves with the animation; the high-resolution mesh and its BVH inside it stay static and reusable. During ray traversal, the system transforms each ray back into the cage’s own reference frame and tests it against the static triangle set inside, so animation cost scales with how complex the cage is, not with how many triangles are inside it.
The demonstration numbers are the kind that normally require more GPU, not a smarter algorithm. Gruen tested the technique on a terrain scene with roughly 25,000 independently animated plants, each responding to a wind-field simulation with locally varying directions, totaling about 2.8 billion triangles before level-of-detail selection. After LOD, the GPU ray-traces around 500 million animated triangles per frame — primary and shadow rays — at more than 60 fps at 1080p, on a single Radeon RX 9070 XT.
| Rendering metric | Conventional BVH | Tetrahedral cage |
|---|---|---|
| Active triangles tested | 500 million | 500 million |
| BVH memory | Up to 80 GB | ~1.7 GB |
| BVH update time | 300+ ms | ~3.3 ms |
Those figures work out to roughly 47 times less BVH memory and a ninety-fold cut in update time. An 80GB acceleration structure exceeds the memory of any gaming GPU on the market; 1.7GB fits on a consumer card with plenty of room left over for shading, AI inference, and everything else competing for the same frame budget.
NVIDIA’s contribution to the same roundup runs in a different direction: folding neural texture, shading, geometry, and rendering stages directly into the RTX pipeline, rather than treating neural inference as a bolt-on step that competes with traditional rendering for GPU time. The two approaches — AMD’s geometric trick and NVIDIA’s pipeline-level neural integration — are solving adjacent problems from opposite ends: one shrinks the cost of exact geometry, the other blurs the line between “exact” and “learned” in the first place.
The anecdote closing out the piece is the one that’ll travel furthest: a developer reportedly had Claude Opus 5.5 write a working dual-GPU path tracer from scratch in about an hour. Paired with the AMD and NVIDIA work, it’s a useful reality check on where the actual bottleneck in real-time graphics sits right now — less “can we render this,” more “how fast can a human or a model get from idea to working renderer at all.”
A demonstration of tetrahedral-cage-driven animated geometry, the technique at the center of AMD’s result.
Related Reading
- AMD, NVIDIA, and AI Rework Rendering — ACM SIGGRAPH Blog
- Ray Tracing Massive Amounts of Animated Geometry Using Tetrahedral Cages — AMD GPUOpen
- 3D Texturing Tutorial: Getting Started With PBR — our tutorial on real-time rendering fundamentals
- MesoCraft Brings Node-Based Procedural Modeling to the Browser — our coverage of another real-time-graphics tool