Microsoft previews DirectX changes to reduce ray tracing memory use
A Microsoft test showed about 90% lower memory use for ray tracing structures, but the company cautioned that games should not expect comparable gains.
Microsoft has outlined two DirectX features in preview that aim to reduce the work and memory needed for ray tracing. Its test of Clustered Geometry showed about 90% lower memory use for acceleration structures, though Microsoft warned that the gains in games would not approach those results.
On a Radeon RX 9070, Microsoft's sample ran at 79 fps with Clustered Geometry, compared with 1.2 fps using conventional rebuilds. The time spent building structures each frame fell to just over 3ms from 788ms. Memory consumption for those structures dropped to 774MB from roughly 7.5GB; those figures do not represent a game's total video memory use.
Clustered Geometry divides models into reusable groups containing up to 256 triangles. Rather than reconstructing an entire object's ray tracing structure when it changes, developers can rearrange existing groups or rebuild a limited number. The companion feature, Partitioned TLAS, divides a scene into regions, limiting updates to areas with movement.
The approach also has a cost: clustered builds require much more temporary scratch memory. Microsoft described the features as the largest DirectX Raytracing change since the API's introduction in 2018.
Nvidia already offers a similar approach through RTX Mega Geometry. The DirectX implementation provides developers with a common route across hardware vendors. Support covers all Nvidia RTX cards, Intel Arc B-series GPUs and recent Core Ultra chips. AMD support currently lists only the Radeon RX 9000 series.
The features still require Windows Developer Mode, and games using them remain some way off. They do not require new hardware, although support depends on the GPU.
