AMD's New Tech Cuts VRAM 47x
AMD published a ray tracing technique called tetrahedral cages on the GPUOpen developer blog. In a demo scene with ~25,000 independently animated plants, the traditional approach consumed up to 80GB of VRAM; with the new technique, only ~1.7GB is needed, and BVH update time drops from 300ms+ per frame to ~3.3ms. A single Radeon RX 9070 XT can deliver 60+ FPS at 1080p with ray tracing enabled.
Why ray-traced grass is so expensive: Ray tracing relies on BVH (Bounding Volume Hierarchy) acceleration structures to determine which triangles a ray hits. Static objects are easy—build the BVH once and reuse it. The pain is animated objects: dense, moving geometry like grass and leaves must rebuild its BVH every frame for every independently animated instance.
In AMD's demo scene, 25,000 plants at the highest detail level represent ~2.8 billion triangles, with ~500 million actually participating in ray tracing each frame. With traditional approaches, VRAM usage peaks at 80GB and BVH updates take 300ms+ per frame. A single frame budget is only 16ms—so BVH updates alone overshoot by nearly 20x. That's why ray-traced vegetation in large open-world games has been practically impossible.
How tetrahedral cages balance the books: The idea is simple—wrap a high-density mesh in a much simpler, deformable "cage." The dense geometry and its BVH stay static—built once and reused; only the cage updates each frame. When a ray enters, the system deforms it according to the cage's deformation, transforming it back into the static reference frame before intersecting with the dense triangles.
This decouples animation cost from triangle count—it now scales with cage complexity instead. In the same scene, VRAM drops from a peak of 80GB to ~1.7GB, and BVH updates drop from 300ms+ to ~3.3ms.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.