Microsoft’s Clustered Geometry shrinks ray tracing memory use from 4.7GB to 267MB in DirectX demo

microsofts clustered geometry shrinks ray tracing memory use from 4 7gb to 267mb in directx demo An NVIDIA GeForce RTX 5080 ran a single demo at 1.4 FPS. With Microsoft's new Clustered Geometry feature turned on, the same card hit 310 FPS, an increase of roughly 221 times.

An NVIDIA GeForce RTX 5080 ran a single demo at 1.4 FPS. With Microsoft’s new Clustered Geometry feature turned on, the same card hit 310 FPS, an increase of roughly 221 times.

The person behind the test cautions that games won’t see a gain anywhere near that size, so the figure is no reason to buy a new GPU.

The memory drop is the bigger story

The frame rate is the eye-catching number, but the memory result may be more important. The acceleration structures the RTX 5080 uses for ray tracing calculations shrank from 4,716 MB to 267 MB, a reduction of about 94%.

Rebuild times fell further still. Reconstructing those structures took 720 milliseconds before the change and only 1 millisecond after.

The tests come from independent researcher Osvaldo Pinali Doederlein, who used a technical demo made specifically for Clustered Geometry. The feature is one of Microsoft’s upgrades to DirectX Raytracing (DXR).

NVIDIA pulls well ahead of AMD here

The AMD Radeon RX 9070 XT ran the same demo with the feature on and reached 91 FPS. Against the RTX 5080’s 310 FPS, NVIDIA’s card comes out about 3.4 times faster in this particular test.

The gap comes from a single synthetic workload. It doesn’t settle how the two cards compare at ray tracing in general.

How Clustered Geometry gets these numbers

When objects move around a scene, the GPU usually has to rebuild large chunks of geometry data again and again. Clustered Geometry lets it reuse smaller structures in place of those rebuilds.

This cuts the compute load sharply, most of all in scenes with complex geometry and many moving objects. Those are the scenes where ray tracing tends to cost the most.

Why your games won’t run 220 times faster

The benchmark is synthetic and was designed to stress the exact operations Clustered Geometry accelerates. Doederlein warned that results like these don’t translate directly to real games, which already rely on several optimization techniques.

The 221x figure measures how much headroom the technique creates in a worst-case scenario. It tells you nothing about the frame rate you’ll get in a released game.

Path tracing is where the gains could matter. If the efficiency carries over beyond the demo, developers could create much more complex path-traced environments without needing a matching increase in processing power or VRAM.

Where you can try it today

Recent NVIDIA drivers already support Clustered Geometry. For compatible Radeon GPUs, AMD offers a dedicated preview driver.

Microsoft still has these DirectX features in testing, available through the Agility SDK 1.722 Preview. Until they exit preview and developers build games that use them, the 267 MB figure is only a demo result.