What Microsoft actually shipped
The DirectX team's announcement, written by engineer Amar Patel, appeared on October 28, 2026. It covers three connected pieces:
- Clustered geometry, which splits objects into small reusable building blocks.
- Partitioned top-level acceleration structures (PTLAS), which let a game update only the parts of a scene that changed.
- A GPU-driven indirect API for building and updating these structures. The CPU describes the work and its bounds, and compute shaders can generate the arguments.
The preview arrives with DirectX 12 Agility SDK 1.722, which introduces Clustered Geometry, PTLAS, and GPU-driven acceleration structure builds. It shipped alongside the retail Agility SDK 1.620. That release is separate and focuses on diagnostics such as DirectX Dump Files. Shader Model 6.10 remains in preview, and Microsoft is targeting an early-December retail release.
Why ray tracing needs this
Real-time ray tracing relies on acceleration structures. These are data structures that let the GPU quickly find which triangles a ray hits. A bottom-level structure (BLAS) describes an object's geometry. A top-level structure (TLAS) describes where all the object instances sit in the scene.
Both get expensive when the world is alive. Animated characters, streamed terrain, procedurally generated surfaces and large instance counts all force rebuilds. Microsoft's spec says the traditional TLAS design requires the whole structure to be rebuilt even if only a few instances change. That is wasteful when most of a scene is static.
Clustered geometry: building with blocks
The unit of work becomes a cluster-level acceleration structure (CLAS). Each CLAS holds up to 256 triangles and 256 unique vertex indices. Developers then assemble a Cluster BLAS from a list of CLAS references. Several Cluster BLAS objects can share the same CLAS, so memory and build work aren't duplicated.
Practical details for developers:
- LOD and streaming: Level-of-detail changes or streamed content can be handled by assembling different sets of existing clusters instead of rebuilding from raw triangles. Meshlet-based pipelines may already produce suitable groupings.
- Deforming meshes: Clustered geometry intentionally has no refit. For animation, cluster templates store topology and non-position data. They are instantiated with current vertex positions to produce a CLAS. The template itself can't be traced.
- Spatial grouping: Clusters should be grouped by proximity to keep trace performance healthy.
- Invalidation: A Cluster BLAS references its source clusters. Changing a referenced CLAS invalidates the dependent Cluster BLAS.
- API differences: Cluster BLAS objects aren't built through the traditional
BuildRaytracingAccelerationStructure()call. Shaders must opt in to traversing them.
Microsoft also argues that cheaper builds could let games trace the same full-detail geometry they rasterize. Today they often trace simplified proxies. Shadows and reflections would then match what's on screen. Gears of War: E-Day is an early example of clustered geometry through Nvidia's RTX Mega Geometry interface. The DirectX version standardizes that idea across vendors.
PTLAS: update only what moved
PTLAS divides the scene into partitions. Spatially coherent instances are grouped, each group gets its own structure, and a second-stage structure combines them. A partition with no changes can keep its built data across frames. To ray-tracing shaders it behaves like an ordinary TLAS.
A few facilities matter for engine programmers:
- Global partition: Frequently moving instances, such as animated characters, can go here. Each behaves, for build and trace purposes, as if it were in its own partition. Microsoft recommends moving them back into a spatial partition once they settle.
- Operation types:
WRITE_INSTANCEwrites full instance records.UPDATE_INSTANCEswaps only the BLAS reference and hit-group contribution, for example on a LOD change.TRANSLATE_PARTITIONsupports origin-relative coordinates in large worlds. - Partition sizing: Microsoft's spec suggests small partitions of roughly 100 to 1,000 instances give most of the build benefit with the least trace penalty. Overlapping partition bounds hurt traversal.
None of this is automatic. Developers have to organize instances and partitions carefully to see the gains.
The benchmark, read carefully
Microsoft's sample runs on an AMD Radeon RX 9070. The scene is what the company calls programmer art. It compares a traditional BLAS rebuild with clustered geometry.
| Metric | Traditional BLAS rebuild | Clustered geometry |
|---|---|---|
| Frame rate | 1.2 FPS | 79 FPS |
| Per-frame GPU build time | 788 ms | 3.11 ms |
| Per-frame GPU ray tracing | 72.6 ms | 9.27 ms |
| Animated acceleration-structure memory | 6,794 MiB | 244 MiB |
| Total acceleration-structure memory | 7,519 MiB | 774 MiB |
Microsoft's own caution matters here. The post says in capital letters not to assume games will see this magnitude of wins, because it's a test playground. Three caveats follow from the figures and the post:
- The memory saving is for acceleration structures only. It is not a cut to total VRAM. Headlines that say "cutting VRAM" are shorthand for a much narrower figure.
- Scratch memory goes up. Static build scratch rises from 1.83 MiB to 526 MiB, and animated build scratch from 8.25 MiB to 179 MiB.
- Tracing can get slower. In a separate PTLAS domino sample, a full TLAS rebuild took 2.88 ms per frame. PTLAS with localized writes took 0.79 ms, but trace time was slightly slower. Microsoft says the trade-off was worth it there.
The 1.2 FPS baseline comes from rebuilding enormous animated structures every frame. That is a stress case, not a typical shipping game. Real titles already work around these costs with simplified proxies and other tricks, so their gains will probably be far smaller.
Hardware and driver support
Microsoft says no new hardware capability is needed beyond existing ray-tracing support. Drivers are another matter. The post says breadth of support depends on vendors implementing and validating drivers for each hardware generation. Support for some older ray-tracing GPUs is still to be determined.
| Vendor | Supported in the preview |
|---|---|
| NVIDIA | All RTX GPUs |
| AMD | Radeon RX 9000 series |
| Intel | Arc B-Series, Core Ultra 200V and Series 3 |
| Qualcomm | Snapdragon X2 platforms, driver support targeted for Q1/Q2 2027 |
Xenospectrum reports that NVIDIA's Game Ready Driver 621.01 supports the preview across GeForce RTX generations, and AMD released a preview-compatible driver for the RX 9000 series. That matches Microsoft's vendor list. Owners of older Radeon cards, such as RX 6000 and 7000, are not on the list for now. Whether they ever will be depends on AMD's engineering decisions.
What developers need to try it
Microsoft's prerequisites for the preview:
- Matching Agility SDK 1.722.0-preview headers and runtime.
- The DirectX Shader Compiler Shader Model 6.10 preview, Patch 2 (v1.10.2605.37).
- A supporting driver.
- Windows Developer Mode, which enables experimental features.
- The preview enabled before device creation. Microsoft's example enables the experimental shader models and raytracing preview features.
- A support check through the raytracing options query, looking at
ClustersAndPTLASSupported, with a fallback to the traditional BLAS/TLAS path.
Microsoft also provides a clustered geometry sample, a PTLAS domino sample contributed by NVIDIA, and PIX preview support. VideoCardz notes the features need compatible drivers and Windows Developer Mode.
What this means for PC gamers
Don't expect your current games to speed up. Because these are new API paths, engines have to adopt them. Digital Foundry says the features should start turning up in games sometime next year. The preview label means the design could still change, and Microsoft has asked developers for feedback before a final release.
The strategic point is vendor-neutral standardization. Until now, clustered geometry in shipped games went through NVIDIA's vendor-specific interface. A common DirectX path could make the technique easier to adopt across GeForce, Radeon and Arc hardware. Qualcomm's X2 commitment suggests Windows on Arm gaming laptops are meant to be included as well.
Bottom line: It's a substantial, well-documented preview with real engineering behind it. The 66x figure is a ceiling set by a synthetic stress test. The realistic expectation is smoother ray tracing in games that animate and stream a lot of detailed geometry. That is most likely to show up first on newer GPUs from 2027 onward.
Update: Independent RX 9070 XT test adds 4K results (November 5, 2026)
According to NoobFeed’s November 5 report, tester Asvaldo found that a Radeon RX 9070 XT reached roughly 28 frames per second at 4K in a geometry-heavy test using clustered geometry—about 17 times the result with the traditional approach.
At 920p, the same report says the test rose from roughly 2 fps with a conventional BLAS setup to approximately 90 fps with clustered geometry. The reported memory figure fell from about 7.5 GB to around 800 MB; that comparison concerns the ray-tracing structures, not the game’s or system’s total memory use.
The additional result gives developers another data point at a higher resolution, but it is still a specialized workload, not a game benchmark. NoobFeed notes that the test does not account for the wider work a game performs, including game logic and physics.
For Windows gamers, the practical implication remains limited: these measurements do not predict a 17-fold or 50-fold increase in existing games. Developers would need to adopt and optimize the preview’s new paths before players could see any benefit.
Update: RTX 5080 test adds new clustered geometry results (November 6, 2026)
According to eGamers.io’s November 6 report, independent researcher Osvaldo Pinali Doederlein tested Microsoft’s clustered geometry demo on an NVIDIA GeForce RTX 5080, reporting 310 frames per second with the feature enabled versus 1.4 fps using the traditional approach.
eGamers.io reports that the demo’s ray-tracing acceleration structures fell from 4,716 MB to 267 MB, while their rebuild time dropped from 720 milliseconds to 1 millisecond. The memory comparison concerns those structures, not total graphics memory use.
The outlet also reports 91 fps for an AMD Radeon RX 9070 XT with clustered geometry enabled in the same demo, compared with the RTX 5080’s 310 fps. That comparison is limited to this synthetic workload and does not establish a general performance gap between the cards.
For Windows gamers, the new figures are not a forecast of gains in existing titles. As eGamers.io notes, the test stresses the operations the feature is designed to accelerate; games already use other optimizations, and developers must adopt the preview technology before players can benefit.
References
- Microsoft’s Clustered Geometry shrinks ray tracing memory use from 4.7GB to 267MB in DirectX demo - eGamers.io eGamers.io · 2026-11-06T19:46:53+00:00
- DirectX Technology Boosts Ray Tracing Performance by 50x on AMD Radeon RX 9070 XT - NoobFeed NoobFeed · 2026-11-05T00:00:00+00:00
- Microsoft Overhauls DirectX Raytracing, Boosting Frame-Rates and Cutting VRAM - Digital Foundry Digital Foundry · 2026-10-30T16:00:00+00:00