AMD detailed a new ray tracing technique called tetrahedral cages. According to a blog post on AMD's GPUOpen developer site, it can cut BVH memory use by roughly 47x while speeding up BVH updates by around 90x. AMD demoed the tech on a large scene with about 25,000 independently animated plants, running at 60+ FPS at 1080p on a Radeon RX 9070 XT.

In a conventional setup, every uniquely animated object needs its own acceleration structure updated every frame, which gets expensive fast in dense scenes. AMD says the same demo would normally require up to 80 GB of GPU memory and over 300 ms per frame for BVH updates using standard methods. With tetrahedral cages, that dropped to about 1.7 GB of memory and roughly 3.3 ms per frame.

The trick is that instead of animating the dense mesh directly, AMD wraps it in a much simpler deformable cage. The dense geometry and its BVH stay static and reusable, and only the cage itself needs to be animated each frame, so update cost scales with the cage's complexity rather than the triangle count.

The tradeoff is losing per-vertex animation control, which is why AMD is positioning it for grass, foliage, and other dense but not individually detailed objects. AMD says the technique can be combined with Microsoft's DirectX ray tracing too. Still, this is only a research showcase in its current state, and there's no word yet on when or if it might show up in games.