Microsoft says it has added clustered geometry, partitioned top-level acceleration structures and GPU-driven acceleration structure builds to DirectX Raytracing as a preview. In its DirectX developer blog post of October 28, part of the Agility SDK 1.722 preview, Microsoft calls it the most significant DXR update since the API debuted in March 2018.
The preview needs no new hardware capability beyond existing raytracing support, but each vendor has to ship a driver for it. NVIDIA covers all RTX GPUs, AMD the Radeon RX 9000 series, and Intel the Arc B-Series and the Core Ultra 200V and Series 3 processors. Qualcomm targets Snapdragon X2 in the first half of 2027, and Microsoft lists support for some older raytracing GPUs as still to be determined. Developers also need the 1.722.0-preview runtime, the Shader Model 6.10 preview compiler and Windows Developer Mode.
Clusters of up to 256 triangles replace whole-mesh rebuilds, and a partitioned top level updates only what moved
A cluster acceleration structure holds an indexed mesh of up to 256 triangles and 256 vertices. An application assembles a bottom-level structure from a list of clusters, so streaming and level-of-detail changes swap clusters instead of rebuilding from triangles. Clustered geometry does not support refits by design. Deforming meshes use cluster templates, which precompute the topology work and are filled with fresh positions each frame.
The partitioned top-level structure applies the same idea to the whole scene. Instances are grouped into partitions, a partition with no changes keeps its built structure, and fast-moving instances can sit in a separate global partition. The third piece, ExecuteIndirectRTASOperations(), lets compute shaders write the build arguments, so the GPU drives the work without a round trip through the CPU.
Microsoft points to Gears of War: E-Day as an early adopter of clustered geometry through NVIDIA’s vendor-specific RTX Mega Geometry API. Putting the feature in DirectX gives other titles one integration path across vendors.
In Microsoft’s Radeon RX 9070 sample, the per-frame build step shrinks from 788 ms to 3.11 ms
The company ran its D3D12 Clustered Geometry sample on a Radeon RX 9070 and compared a traditional bottom-level rebuild with clustered geometry. Frame rate went from 1.2 FPS to 79 FPS, per-frame GPU build time from 788 ms to 3.11 ms, and per-frame GPU raytracing from 72.6 ms to 9.27 ms. Microsoft calls the scene a test playground, describes it as one configuration, and warns in capitals that games should not assume this magnitude of wins.
The 89.7% memory saving comes mostly from animated geometry, while build scratch space grows
Microsoft says its table puts the headline memory figure at a total of 7,519 MiB falling to 774 MiB. Split out, the static acceleration structure goes from 723 MiB to 527 MiB, a 27.0% saving, while the animated one goes from 6,794 MiB to 244 MiB, a 96.4% saving. The total may therefore depend on how large the traditional rebuild path’s animated baseline is, so a real game with a smaller animated share could land lower.
Scratch memory moves the other way. Static build scratch rises from 1.83 MiB to 526 MiB, and animated build scratch from 8.25 MiB to 179 MiB. Microsoft labels both as a higher scratch requirement.
The partitioned top level shows a similar trade in a second sample, a world of falling dominoes. A full TLAS rebuild each frame takes 2.88 ms, while the partitioned structure with 4702 localized writes takes 0.79 ms. Trace time is a little slower with partitions, and Microsoft judges the trade worth it in that scene.
Everything above comes from Microsoft’s own samples. The post gives no retail date and asks developers for feedback on Discord before the design is finalized, and NVIDIA’s Zorah demo scene is noted as untested on other vendors’ GPUs.