· Mick
AMD cuts ray-tracing BVH memory in a demo: what the 98% figure means
AMD is showing a way to keep billions of animated triangles available to ray tracing without rebuilding all of that geometry every frame. The measured gain is enormous, but it needs context before anyone turns it into a claim about a game’s total VRAM use.
Analysis of AMD technical publications. GPUStores has not reproduced the demo or benchmarked the RX 9070 XT.
AMD published the demonstration on September 17, 2026. It contains about 25,000 independently animated plants. They represent 2.8 billion triangles at maximum LOD; after LOD selection, roughly 500 million animated triangles are traversed per frame with primary and shadow rays.
The numbers: 1.7GB and 3.3ms, in this specific test
On a Radeon RX 9070 XT at 1080p, AMD reports more than 60 FPS. The tetrahedral cages use about 1.7GB of BVH memory, and all BVH updates take roughly 3.3ms per frame. A conventional dense-triangle BLAS implementation would require up to 80GB and more than 300ms for the same independently animated scene.
Moving from 80GB to 1.7GB works out to about 97.9%, commonly rounded to 98%. That percentage applies to acceleration-structure memory for this scene. It is not a 98% cut to a game’s total VRAM use, and it is not a general GPU comparison.
The cage moves; dense geometry stays reusable
A conventional scene needs unique acceleration-structure state for each deforming object. Here, a dense mesh is enclosed in a much smaller tetrahedral cage. Only the cage is animated. Detailed geometry and its BVH stay static, while rays are transformed back into that reference space before intersection.
The cost therefore follows cage complexity instead of the full triangle count. AMD’s technical article describes reusable mini-BLAS data and how the approach can work with DirectX Raytracing.
The trade-off: less control over each vertex
The efficiency has a clear limit: developers no longer control every vertex in the dense mesh independently. AMD presents cages as a fit for foliage, grass, crowds and some creatures. Cluster-level acceleration structures may still be preferable when an animation needs the best local precision.
This is not a switch in AMD Software: Adrenalin and it is not a new version of FSR. A studio must integrate the technique into its engine and select geometry that tolerates this kind of deformation.
What this changes for you today
This demo alone is not a reason to buy a GPU. No game, commercial engine adoption or deployment schedule has been confirmed. The test uses one RX 9070 XT at 1080p; it does not establish results on other Radeons, GeForce cards, 1440p or 4K.
The potential is for future games: more moving foliage or crowds under ray tracing without overwhelming BVH memory and update budgets. For a current purchase, keep judging total capacity with our VRAM guide and known performance in our RX 9070 XT comparison. Our ray-tracing technology tracker also separates shipping features from announcements.