Microsoft has put a big change to DirectX Raytracing into preview, and its own test scene shows what's at stake: ray-tracing memory down from about 7.5 GB to 774 MB on an AMD Radeon RX 9070. Microsoft also attaches a loud warning that games shouldn't expect results like that.
The features are called Clustered Geometry and Partitioned Top-Level Acceleration Structures (PTLAS). Microsoft introduced them on its DirectX Developer Blog on 28 October as part of Agility SDK 1.722 preview, and called it "the most significant update since its debut in 2018" for DXR. PC Gamer covered the preview on 2 November, describing it as bringing tech very similar to Nvidia's RTX Mega Geometry to every other GPU vendor.
What actually changes
Ray tracing works by testing rays against acceleration structures, data structures that describe the scene's geometry, and rebuilding them is expensive. Microsoft says animation, streaming and generated geometry can repeatedly force costly rebuilds of large bottom-level structures, while updating the top-level structure for a big world can be costly even when only a few objects have moved.
Clustered Geometry breaks that work into small reusable clusters. Microsoft's blog puts each cluster at up to 256 triangles and 256 vertices. Partitioned TLAS splits the scene into regions, so the cost of an update tracks what changed rather than the whole world. Microsoft says the two can be used separately or together, and that more efficient builds could make it practical to ray trace the same full-detail geometry used for rasterisation, so ray-traced shadows and reflections match the visible scene.
Microsoft also says the features need no new hardware capabilities beyond existing ray-tracing support. The blog lists supported hardware and drivers as all Nvidia RTX GPUs, AMD Radeon RX 9000 series, and Intel Arc B-Series plus Core Ultra 200V and Series 3 processors. Qualcomm says it is targeting driver support for Snapdragon X2 in the first half of 2027. Support for some older ray-tracing GPUs is still to be determined, according to Microsoft.
The 90% figure, and the caveat in capitals
The headline numbers come from Microsoft's sample running on an RX 9070. Against a traditional rebuild, Microsoft's table shows:
- total acceleration-structure memory falling from 7,519 MiB to 774 MiB, or 89.7% less
- frame rate rising from 1.2 fps to 79 fps
- per-frame build time falling from 788 ms to 3.11 ms
Microsoft is blunt about what that means: "DO NOT ASSUME GAMES WILL SEE THIS MAGNITUDE OF WINS, THIS IS JUST A TEST PLAYGROUND." The same table shows build scratch memory going up sharply with clusters, and PC Gamer flags that extra temporary memory as a new problem for developers to manage.
Microsoft names Gears of War: E-Day as an early game already using clustered geometry through Nvidia's own RTX Mega Geometry API, and says standardising it in DirectX gives developers a common path across supporting hardware.
Not something you can switch on
This is a developer preview. Microsoft says it needs matching Agility SDK preview headers and runtime, a supporting driver and Windows Developer Mode, and that feedback will shape the final design before a retail release. PC Gamer notes there's no word on when players will see it. Games will have to adopt it first.
Mason's take: The 774 MB number will travel, and it shouldn't travel alone. Microsoft wrote its warning in capitals for a reason. What matters more is the shape of the change. Rebuilding less of the scene each frame is how ray tracing gets cheaper on cards people already own, and putting it in DirectX rather than one vendor's API is what gets it into more games. With memory and GPU prices where they are, that's the part worth watching.
Reporting notes
