Microsoft previews DirectX ray-tracing features that cut memory by about 90 percent

Microsoft said in late October 2026 on its developer blog that Clustered Geometry and Partitioned TLAS cut ray-tracing memory by about 90 percent on a Radeon RX 9070 in a preview sample.

Microsoft described the pair as the biggest change to DirectX Raytracing since the API arrived in 2018. Acceleration structures work like a table of contents for scene geometry, letting the GPU discard entire sections during ray-triangle intersection tests instead of testing every ray against every triangle. Microsoft positioned the update as infrastructure for large busy game worlds rather than a visual feature.

Clustered Geometry splits object meshes into reusable clusters, each with no more than 256 triangles and 256 vertices, letting engines assemble structures from existing clusters instead of rebuilding from individual triangles each frame. The compact vertex format COMPRESSED1 records only distances to a shared cluster anchor point instead of three full floating-point values per vertex.

Partitioned TLAS splits the game world into spatial regions, updating only the partitions containing changed objects and leaving static partitions intact across frames. The PTLAS operations include WRITE_INSTANCE for moving objects, UPDATE_INSTANCE for swapping referenced bottom-level structures such as detail changes, and TRANSLATE_PARTITION for shifting an entire region without touching every instance individually. Constantly moving objects go into a global partition where the engine treats each as its own instance and moves them back once they come to rest.

Microsoft said in a preview sample that ray-tracing acceleration structure memory on a Radeon RX 9070 dropped to 774 MiB from roughly 7.3 GiB or 7.5 GB, a cut of about 90 percent. A Microsoft benchmark measured by Igor’s Lab on Radeon RX 9070 showed conventional structures at about 12,728 MiB versus 774 MiB for the new features, a cut of about 93.9 percent. On a Radeon RX 9070, Microsoft also saw animated acceleration structure memory fall to 2,448 MiB from about 11,359 MiB.

In Microsoft’s Clustered Geometry sample on a Radeon RX 9070, frame rate rose from 1.2 fps to 79 fps while per-frame build time fell from 788 ms to about 3 ms. Igor’s Lab recorded GPU build time at 3.06 ms per frame down from 482 ms, and ray tracing time at 8.47 ms down from 73.8 ms, yielding 84 fps versus 1.8 fps.

Microsoft warned in all capital letters that games should not expect performance wins close to the sample sizes. The sample rebuilt animated objects every frame, something real games avoid through existing optimizations, and Microsoft said the figures show the technique’s upper limit instead of typical gameplay performance. Temporary working memory during builds grows; Igor’s Lab pointed to static structures rising from 1.78 MiB to 526.44 MiB in one example.

Microsoft said the features work with existing ray tracing hardware but need compatible drivers, and each vendor decides actual support. Nvidia supports every GeForce RTX GPU, while Club386 names only the GeForce RTX 50 series and TechSpot, TweakTown, and Basic Tutorials say every RTX generation or all GeForce RTX GPUs will support the features after driver updates. AMD supports only the Radeon RX 9000 series for now, leaving out Radeon RX 7000 and RX 6000 series even though both support ray tracing.

Intel supports Arc B-series GPUs and recent Core Ultra chips such as the 200V series and Series 3, while Arc A-series is not listed. Basic Tutorials cites Battlemage cards such as the Acer Nitro Intel Arc B570 and B580. Qualcomm support is expected later, with Basic Tutorials pointing to Snapdragon X2 in the first half of 2027 and HotHardware saying Adreno support is coming next year.

Current titles will not run faster with this update; developers must explicitly integrate Clustered Geometry and PTLAS into their engines. Microsoft made DirectX 12 Agility SDK 1.722 Preview available on Oct. 28, 2026, alongside retail SDK 1.620, and the preview capabilities are included. The preview requires Windows Developer Mode, so it is not intended for players at this time.

Shader Model 6.10 remains in preview, and Microsoft plans a retail release for it alongside Agility SDK 1.621 in early December 2026 without setting a retail date for the ray-tracing features themselves. Microsoft is gathering feedback on the preview through Discord and plans to move to the regular version as soon as possible, with no final release date announced. Basic Tutorials assessed that large engines using streamed detail geometry and path-traced titles are most likely to adopt the technology first, and widespread use could take at least a year.

Nvidia shipped RTX Mega Geometry in 2025 as a vendor-specific feature, and it appeared in Alan Wake 2 and Gears of War: E-Day. TechSpot reports that RTX Mega Geometry boosted both frame rates and image quality in Alan Wake 2, and that Gears of War: E-Day used RTX Mega Geometry to ray trace Unreal Engine 5‘s full-detail Nanite geometry rather than static low-poly proxies.

To save computation, many games currently use simplified low-detail replacement geometry for ray tracing, though the results do not exactly match the visible scene, especially with detailed streamed geometry. Microsoft is bringing cluster-based ray tracing into DirectX, so developers can use one cross-vendor path instead of maintaining an Nvidia-only solution. To store batch build arguments in graphics memory, GPU-driven acceleration structure builds use ExecuteIndirectRTASOperations(), letting compute shaders generate them while the CPU only sets up the framework.

Retail Agility SDK 1.620 includes Partial Graphics Programs for shader reuse and DirectX Dump Files for standardized GPU crash reports in finished games, initially limited to Radeon RX 7000 and RX 9000 GPUs. The DirectX specification defines D3D12_RAYTRACING_TIER_2_0 as a new tier bundling Clustered Geometry, PTLAS, and all previous DXR features.