AMD’s Tetrahedral Cages Slash Ray Tracing Memory Usage From 80GB to Just 1.7GB

Ray tracing has always demanded a brutal trade-off: the more realistic the lighting and shadows, the more video memory a graphics card burns just keeping track of where everything is. AMD’s research division may have just cracked that problem wide open, and the numbers involved are almost hard to believe.

At High-Performance Graphics 2026, AMD engineers unveiled a technique called tetrahedral cages, a new way of handling Bounding Volume Hierarchies (BVH) — the internal “maps” a GPU uses to know where every object in a scene is located so it can trace rays correctly. The work was significant enough to win third place for the Wolfgang Straßer Best Paper Award at the conference.

Why BVH Memory Is Such a Problem

A video game is essentially a rapid slideshow, rendering 60 or more frames every second. For ray-traced lighting and shadows to look convincing, the GPU must constantly rebuild its BVH data every time something on screen moves — a swaying tree, a running crowd, or drifting grass. Doing this for thousands of individually animated objects at once has traditionally required enormous amounts of VRAM and processing time, which is why fully dynamic, dense ray-traced environments have remained mostly out of reach for real-time games.

Putting Objects in an Invisible Box

AMD’s fix is elegantly simple in concept. Instead of updating the detailed BVH data for every moving object every frame, the technique wraps each object in a lightweight, simplified “cage” — a tetrahedral mesh that deforms and moves in its place. The dense, complex geometry inside stays static and reusable; only the simple cage needs to be tracked and animated. When a ray needs to intersect with the object, it is transformed back into the object’s static reference frame, so the heavy geometry and its acceleration structure never actually have to move.

Because animation cost now scales with the simplicity of the cage rather than the complexity of the object itself, this approach is particularly suited to foliage, grass, crowds, and distant characters — scenes with huge numbers of moving objects that don’t need fine, vertex-level deformation.

The Demo Numbers

AMD demonstrated the technique using a scene packed with roughly 25,000 independently animated plants, representing about 2.8 billion triangles at maximum level of detail, with around 500 million animated triangles ray-traced per frame across primary and shadow rays. Running on a single Radeon RX 9070 XT, the results were dramatic.

That works out to roughly 47 times less memory consumption and a 90 times faster BVH update speed, while still holding a smooth 60-plus frames per second — a scenario that would previously have been completely unworkable on consumer hardware.

Still a Research Project — For Now

AMD has been clear that tetrahedral cages remain in the research phase and are not yet part of a shipping driver, game, or SDK feature. The company says it is developing DXR sample code and a C++ library so game developers can begin experimenting with the technique, which suggests real-world adoption could still be some way off. Even so, this is exactly the kind of foundational research that tends to quietly reshape what’s possible in future graphics cards and game engines, much like earlier BVH compression work eventually fed into mainstream ray-tracing features.

For gamers, the appeal is obvious: techniques like this could eventually allow far denser, more dynamic ray-traced worlds — think entire forests or city crowds rendered with full lighting accuracy — without demanding GPUs stacked with impractical amounts of VRAM. AMD’s own Radeon RX 9070 XT, a mid-to-high-end card rather than a flagship, achieving this result also hints that such visual fidelity may not remain locked to only the most expensive hardware. Readers tracking the next generation of AMD Radeon graphics cards on TechnoSports will want to keep an eye on how this research evolves.

Source: AMD GPUOpen