Microsoft's DirectX Raytracing preview cuts build times and VRAM use sharply

Here is what is physically different under the hood: Microsoft’s DirectX Raytracing preview changes how the acceleration structures are built, and that is where a lot of ray tracing time goes. These are the data trees the GPU uses to find what a ray hits. The company’s engineer calls it the most significant DXR update since the standard debuted in 2018.¹

The update rests on two ideas. Clustered geometry splits a model into small reusable pieces of up to 256 triangles and 256 vertices each, instead of treating it as one huge object.¹² If part of the model changes, developers can reuse existing chunks or rebuild only the small sections that changed.³ Think of it as building from blocks instead of carving from a single plank of wood.

The second idea, partitioned top-level acceleration structures (PTLAS), applies the same logic to the whole game world. When something moves in one area, only that region’s structures are recalculated and the rest of the scene stays as it was.³² Both features can be used separately or together, and builds can be driven from the GPU itself.¹

Now the numbers, with the test setup attached. In a Microsoft sample running on a Radeon RX 9070, switching from traditional bottom-level structures to clustered geometry took one scene from 1.2 fps to 79 fps, roughly 66x faster.³² Build time fell from 788 ms to 3.11 ms, and ray tracing time dropped from 72.6 ms to 9.27 ms.³² Memory for acceleration structures shrank from about 7.3 GB to 774 MB.³²

I would not read those as what your games will do. The sample is a purpose-built test scene, and Microsoft’s own slide warns not to assume games will see wins of that size.³ There is also a trade-off: scratch memory, the temporary workspace used during builds, rose substantially in the test.²

One practical upside is that detail can stop being sacrificed. Cheaper builds make it plausible to trace against the same full-detail geometry used for drawing the screen, rather than simplified stand-ins, so shadows and reflections match what you see.¹ Nvidia’s RTX Mega Geometry, used in Gears of War: E-Day, is an early vendor-specific take on the same idea. This is a standardized version meant to work across vendors.³¹

No new hardware capability is required, but drivers are. Nvidia says all RTX GPUs support the new version, while AMD support starts with Radeon RX 9000 and Intel’s covers B-Series GPUs and recent Core Ultra chips.³² Qualcomm’s X2 is promised by the middle of next year.³ Support for older ray-tracing GPUs is undecided, since implementation work may turn up performance problems.¹

The preview ships in the DirectX 12 Agility SDK 1.722 preview now, and games using it could start appearing next year.³

Our read: A 66x headline number is the kind of “up to” figure I distrust, and the extra scratch memory shows nothing here is free. But the memory result is the interesting part. Cutting acceleration structure storage by roughly 90 percent in a test is the sort of boring engineering win that frees VRAM for everything else. Whether real games get even a fraction of it depends on drivers and on developers adopting it.

On the bench: if you own an RTX card, a Radeon RX 9000 or a recent Intel Arc B-Series, you are covered on paper, but nothing changes until games ship with this support. Older ray-tracing cards are the ones to watch.

Footnotes

1. [DirectX Developer Blog, “PREVIEW) Raytracing: Clustered Geometry, Partitioned TLAS, Indirect Acceleration Structure Builds”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Wccftech, “Microsoft Unveils Biggest DirectX Raytracing Upgrade Since 2018, Bringing More Efficient Ray Tracing To NVIDIA, AMD & Intel GPUs”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. digitalfoundry.net, “Microsoft overhauls DirectX Raytracing for big performance gains”. Retrieved October 31, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰