Microsoft has published a preview of the two largest additions to DirectX Raytracing since the technology shipped in 2018: Clustered Geometry and Partitioned Top-Level Acceleration Structures, or PTLAS. Both arrive in the Agility SDK 1.722 preview, and the DirectX team calls the pair “the most significant update since its debut in 2018.”
The release follows a run of DXR updates. DirectX Raytracing debuted in March 2018, gained inline ray tracing and GPU-driven ray dispatch with Tier 1.1 in November 2019, added payload access qualifiers in Shader Model 6.6 in April 2021, and picked up Shader Execution Reordering and Opacity Micromaps in Shader Model 6.9 earlier this year. Microsoft frames this preview as the next step above all of them.
The two features attack the same cost from opposite ends of a scene. Clustered Geometry breaks the work of building bottom-level acceleration structures into small, reusable triangle clusters, so games can ray trace the same full-detail geometry they already render rather than a simplified stand-in. PTLAS splits the top-level structure into partitions, so a change in one part of a large world no longer forces a rebuild of the entire scene. Neither feature requires new hardware beyond existing ray tracing support.
What Clustered Geometry changes
Acceleration structures are the spatial indexes a GPU walks when it traces a ray against a scene. Building them is expensive, which is why ray-traced games often trace a low-detail proxy instead of the mesh the player sees. Clustered Geometry shrinks the unit of work: a Cluster Level Acceleration Structure, or CLAS, holds an indexed mesh of up to 256 triangles and 256 vertices. A game assembles a Cluster BLAS from a list of these small clusters, and multiple Cluster BLAS can share the same CLAS. Each cluster carries a 32-bit ClusterID that hit shaders retrieve alongside a local PrimitiveIndex, which lets a game look up normals and materials for the exact triangle a ray struck.
The payoff is reusability. Streaming, level-of-detail transitions, and procedurally generated geometry can assemble different sets of existing clusters instead of rebuilding a full structure from raw triangles. For animated meshes, cluster templates precompute the position-independent work once, then instantiate it each frame with fresh vertex positions. Microsoft notes that once builds get cheap enough, it becomes practical to ray trace full-detail geometry, so shadows and reflections match what the player sees.
The build path is GPU-driven end to end. A game queries sizing and scratch requirements with GetRTASOperationPrebuildInfo, then submits batch operations through ExecuteIndirectRTASOperations using argument arrays that compute shaders wrote into GPU memory. Dependent stages, such as cluster builds feeding a Cluster BLAS that then feeds a PTLAS, run in separate calls with barriers between them. Shader Model 6.10 also adds TriangleObjectPositions intrinsics, so a hit shader can recover a triangle’s object-space vertices from the acceleration structure itself instead of fetching them from a separate vertex buffer.
Clusters also support a compressed vertex path. The COMPRESSED1 position encoding stores a shared anchor and a power-of-two scale per cluster, then packs integer offsets from that anchor instead of three full floating-point values per vertex. It shrinks the memory and bandwidth of the geometry that gets traced, which is where Microsoft’s sample shows its biggest savings.
Full path tracing already landed this year in games like the Witcher 3 Remastered, which paired path tracing with DLSS 4.5; Clustered Geometry is the plumbing that could make tracing the real, un-proxied mesh cheap enough to become the default.
Partitioned TLAS updates only what moves
At the scene level, updating a top-level acceleration structure is costly for big worlds even when only a few objects moved. PTLAS assigns instances to persistent partitions, and a partition that received no changes keeps its already-built structure across frames. Update cost scales with what changed instead of the size of the whole world. Frequently moving objects can sit in a separate global partition so they stop disturbing spatial groups, then return once they settle down.
PTLAS adds operation types for the changes open worlds need. WRITE_INSTANCE writes complete instance records, UPDATE_INSTANCE swaps a referenced BLAS for a level-of-detail change without touching the transform, and TRANSLATE_PARTITION moves a whole partition in origin-relative large-world coordinates without rewriting every instance. The update path is fully GPU-driven: compute shaders generate the build arguments, and applications submit them in batches through ExecuteIndirectRTASOperations with no CPU round-trip. Sparse changes write only the affected instance records; everything unwritten stays as it was.
The numbers behind the preview
Microsoft published its own measurements from a sample running on an AMD Radeon RX 9070, comparing a traditional per-frame BLAS rebuild against Clustered Geometry. Per-frame acceleration-structure build time fell from 788 milliseconds to 3.11 milliseconds, and per-frame ray tracing dropped from 72.6 milliseconds to 9.27 milliseconds. Memory moved further: animated acceleration-structure memory fell from 6,794 MiB to 244 MiB, and total acceleration-structure memory from 7,519 MiB to 774 MiB.
The company tempers those figures in the same post. The sample is a test playground, and Microsoft writes in all caps that developers should not assume games will see this magnitude of wins. Build scratch memory rises in the clustered path, from 1.83 MiB to 526 MiB for static geometry. A separate domino sample from NVIDIA shows the PTLAS benefit more modestly: a full top-level rebuild took 2.88 milliseconds, while 4,702 localized writes through PTLAS took 0.79 milliseconds.
Who can try it, and what still has to happen
The preview needs matching Agility SDK preview headers and runtime, a supporting driver, and Windows Developer Mode. Clustered Geometry’s shader intrinsics use Shader Model 6.10. NVIDIA said it will “fully support this update to DXR on all RTX GPUs,” and its engineering lead framed the features as unlocking “new algorithms including path tracing of hierarchical LOD streaming solutions.” AMD listed Radeon RX 9000 series cards, the same vendor that separately extended FSR 4.1 upscaling to older Radeon RX 6000 GPUs earlier this month, and pointed to better “scalability for streaming content, level-of-detail transitions, animation, and large open-world experiences.” Intel listed Arc B-Series graphics and Core Ultra 200V and Series 3 processors, and Qualcomm is targeting Snapdragon X2 platforms in the first half of 2027. Support for some older ray tracing GPUs is still undecided, because implementation work may uncover performance problems.
Microsoft shipped the preview with the tooling developers need to adopt it. Samples for clustered geometry and partitioned TLAS are available in the DirectX-Graphics-Samples repository, NVIDIA contributed its own RTX Mega Geometry sample, and PIX gained preview support for inspecting clusters and partitions. The WARP software renderer is available through NuGet for machines without a ray tracing GPU.
Gears of War: E-Day is an early example of a title already using clustered geometry through NVIDIA’s vendor-specific RTX Mega Geometry API. Standardizing the capability in DirectX gives every vendor one integration path. Because this is a preview, the API surface can still change before a retail release, and Microsoft is asking for feedback through its Discord before it either refines the design or pushes the features to retail.
