The figures come from an AMD GPUOpen post dated September 17, following the company’s broader July research write-up by AMD Fellow Holger Gruen and colleagues. PC Games Hardware independently reported the earlier combined demonstration at approximately 585 million animated triangles and 60 fps on the same Radeon RX 9070 XT. The difference is meaningful rather than contradictory: AMD’s newer terrain video isolates the plant scene at about 500 million triangles after LOD selection, while the earlier figure describes a combined scene that also includes other assets.
For Windows PC gamers, graphics programmers, and DirectX developers, the finding is a reminder that geometry scale in ray tracing is increasingly limited by the cost of maintaining moving acceleration structures—not only by shader throughput or raw ray-intersection hardware. AMD has demonstrated a potentially useful answer for vegetation, grass, crowds, and similarly approximate motion. It has not announced a Radeon driver feature, a DirectX API feature, an Unreal Engine plug-in, or a game ready to turn on this capability.
The 80 GB comparison is a feasibility test, not a VRAM measurement
AMD’s most arresting number needs careful reading. The Radeon RX 9070 XT test platform does not somehow provide 80 GB of usable GPU memory. AMD is saying a conventional dense-triangle bottom-level acceleration structure, or BLAS, for the same set of independently deforming plant instances could require up to 80 GB and take more than 300 ms to update every frame.
That makes the ordinary approach unworkable on a consumer graphics card before the renderer has budgeted memory for textures, buffers, render targets, shaders, or the rest of the game. A 300 ms BVH update also limits rendering to well below four frames per second even if every other part of the frame were free.
AMD’s cage-based path uses around 1.7 GB for its BVHs and spends roughly 3.3 ms updating them. Those two results address separate constraints. The memory saving determines whether the scene can reside on the GPU at all; the update-time saving determines whether it can coexist with a 16.7 ms frame target for 60 fps.
The comparison is still vendor-provided and belongs to AMD’s chosen foliage scene, rendering setup, and implementation. It should not be read as a universal claim that every animated ray-traced game will gain 90 times the performance or use 47 times less memory. But it does establish why the technique is interesting: it converts a conventional workload that exceeds both a consumer GPU’s memory capacity and its frame-time budget into one that has room to run.
Animate the proxy, reuse the expensive geometry
A BLAS is the lower-level hierarchy hardware ray-tracing pipelines use to avoid testing every ray against every triangle. It is an essential shortcut, but animation makes it expensive. When a dense object deforms, its vertex positions change; a renderer must refit or rebuild an acceleration structure that represents the new spatial arrangement.
With thousands of plants responding differently to a local wind field, conventional instancing loses much of its usual advantage. Each uniquely deformed plant effectively needs its own current geometric state and associated acceleration data. Multiplying dense meshes by thousands is not the hard part; updating their individual BVHs every frame is.
AMD’s technique puts a much coarser deformable tetrahedral cage around each dense object. Before rendering, the detailed original mesh is split into pieces associated with the tetrahedra, and static mini-BLASes are built for those pieces. At runtime, the renderer animates the lightweight cage rather than the enormous collection of source triangles.
When a ray reaches one of those moving tetrahedra, the renderer maps the ray back into the object’s unanimated reference space. It can then intersect the static high-detail geometry and its reusable acceleration structures. The dense mesh remains detailed for ray intersection, but it no longer drives the frame-by-frame animation and BVH-update cost.
That is the paper’s actual contribution: the cost of animation becomes tied much more closely to cage complexity than triangle density. It is a data-structure and representation change, not a new kind of Radeon ray-tracing core and not a replacement for normal triangle rendering.
The demo is large, but it is deliberately narrow
AMD’s September terrain scene begins with plants totaling about 2.8 billion triangles at their highest detail levels. Level-of-detail selection reduces the active ray-traced total to roughly 500 million animated triangles per frame. The rays are primary and shadow rays, and AMD explicitly says the video concentrates on massive animated geometry rather than advanced lighting.
That qualifier matters. The result is impressive, but it is not equivalent to a full path-traced open-world game at 1080p with multiple diffuse bounces, glossy reflections, transparency, volumetrics, temporal denoising, AI upscaling, post-processing, and gameplay running at the same time. Nor does it show what frame rate a shipping game would reach after CPU simulation, streaming, materials, character systems, and all of the other work surrounding the renderer.
AMD also gives no public apples-to-apples image-quality metric for the cage approximation in this specific demo, no source package yet, and no announced game or engine partner using it. The company says it is working on DirectX Raytracing samples and a header-only C++ library for constructing high-quality cages, but has not provided a release date.
Those omissions do not undercut the core result. They define it. This is graphics research with a well-scoped demonstration, rather than a promise that a future Adrenalin driver will make existing games render forests at a fraction of their present cost.
Foliage is the obvious target, and the trade-off is visible control
Tetrahedral cages deliberately surrender some animation precision. Their deformation is piecewise linear: vertices move according to the cage around them instead of receiving fully independent, exact deformation. A finer cage can track an animation more closely, but also costs more memory and update time; a coarser cage is cheaper but may produce less faithful motion.
For wind-blown vegetation, grass patches, background crowds, distant creatures, or animation LOD, that is often an appropriate bargain. These are objects whose motion is visually important in aggregate but whose individual vertices are rarely under close inspection. A forest can look wrong if every tree moves identically, yet it does not necessarily require film-quality leaf deformation to look convincing during play.
The technique is much less suitable where a viewer will notice small facial movements, sharp folds, contact deformation, topology changes, tearing, destruction, or precise skinning. AMD explicitly identifies topology-changing animation, fine-scale movement, and sharp deformations as poor fits. Developers would need conventional dynamic BLAS updates, cluster-level acceleration structures, or other techniques for those assets.
This points toward a hybrid renderer rather than an all-or-nothing change. A game could use cages for a distant tree canopy and crowd fill while reserving higher-fidelity dynamic geometry processing for hero characters, vehicles, weapons, and destructible objects. AMD says the technique can coexist with cluster-level acceleration structures and can be combined with Microsoft’s DirectX Raytracing Functional Spec, Part 2, including partitioned top-level acceleration structures.
DirectX support means a possible route, not deployment
AMD’s reference to DXR is encouraging for Windows developers because it suggests the idea can be implemented within the DirectX ray-tracing pipeline rather than requiring a proprietary game-facing graphics API. The technical work, however, belongs primarily in an engine’s content pipeline and ray-tracing implementation: cage generation, association of mesh pieces with tetrahedra, ray transformations, handling of edge cases, and choices about visual error all need to be solved by the developer.
That is a higher bar than dropping in a post-processing library. The eventual header-only C++ library could reduce the asset-processing burden, but it will not automatically make an existing DXR title compatible. Engine teams would have to decide which asset categories merit the extra representation and test whether cage artifacts remain acceptable across animation, camera distance, lighting, and gameplay conditions.
The High-Performance Graphics 2026 awards list places Ray Tracing Massive Amounts of Animated Geometry in a tie for third for the Wolfgang Straßer Best Paper Award. That external recognition supports the research significance, though it is not evidence of production readiness. The companion fact worth watching is whether AMD publishes its promised DXR samples and cage-generation library with enough detail for independent developers to reproduce the memory and update-time results.
For now, the practical takeaway is restrained but substantial: AMD has shown that unique animation, rather than triangle count alone, is a major ray-tracing memory problem, and that a reusable static representation can sharply reduce it when the animation can tolerate approximation. The next concrete milestone is software, not another benchmark—working DXR samples that let Windows graphics developers test whether the 1.7 GB result survives outside AMD’s controlled foliage demo.