Thursday, October 29th 2026

DirectX Raytracing Gets Clustered Geometry for Higher FPS and Up to 90% VRAM Savings

Microsoft has released one of the most significant updates to its DirectX Raytracing (DXR) since its debut in 2018. This update officially introduces clustered geometry and partitioned top-level acceleration structure (PTLAS). With clustered geometry, Microsoft envisions scenes where all animated and generated geometry is broken down into simple, reusable clusters of triangles. Typically, scenes repeatedly build large bottom-level acceleration structures, which incur additional compute costs. By using clustered geometry, ray tracing can utilize full detailed geometry used for standard rasterization instead of proxies, ensuring that ray-traced shadows and reflections match the currently visible scene.

In open-world games, using a top-level acceleration structure (TLAS) can lead to massive VRAM and compute demands. To address this, DXR now introduces partitioned TLAS (PTLAS), allowing ray tracing to focus on changes in a specific region while retaining acceleration-structure data elsewhere. Clustered geometry and PTLAS can be used together or separately, with no limitations. Below image is NVIDIA's sample of clustered geometry in action, visualized in three stages: default, clusters, and triangles applied.
Benchmarks follow.

Benchmarking clustered geometry on the AMD Radeon RX 9070 GPU shows that the improvements in FPS are significant, and VRAM savings are substantial. Frame rates can be up to 66 times higher, though in real games, we can generally expect a few percentage points of improvement. However, VRAM consumption has been reduced by nearly 90%. All current-generation hardware supports these features, with day-one support from AMD, Intel, NVIDIA, and Qualcomm for their graphics solutions.
Sources: Microsoft, Thanks TPU Forum member TumbleGeorge for the tip!
Add your own comment

11 Comments on DirectX Raytracing Gets Clustered Geometry for Higher FPS and Up to 90% VRAM Savings

#1
AGlezB
We use averages and 1% lows to measure variables because "up to" usually means "we crafted a beautiful tech demo producing results you'll never see IRL". Great for marketing tho. :laugh:
Posted on Reply
#2
jdrch
AGlezBWe use averages and 1% lows to measure variables because "up to" usually means "we crafted a beautiful tech demo producing results you'll never see IRL". Great for marketing tho. :laugh:
VRAM usage is the key benefit here. One benefit of the DDR5 shortage is it should finally put an end to the largely pointless framerate chasing post-COVID.
Posted on Reply
#3
AGlezB
jdrchVRAM usage is the key benefit here. One benefit of the DDR5 shortage is it should finally put an end to the largely pointless framerate chasing post-COVID.
My point is that the "up to" value is the maximum under optimal/lab conditions, not the value you'll see with any regularity during normal use. It doesn't matter what is it that your measuring, be framerate, VRAM usage, CPU/GPU usage, etc. and is not limited to computers either.
Posted on Reply
#5
PierreJG
The research I did was that, for example, the RTX 3060 will have day-one support; the same goes for the RX 9xxx GPUs, but older AMD cards will not.
They call it hurdles; I just call it AMD and Microsoft being AMD and Microsoft. This could actually benefit both the RX 6xxx series and the 7xxx series.
Posted on Reply
#6
Sspacenaxe
PierreJGThe research I did was that, for example, the RTX 3060 will have day-one support; the same goes for the RX 9xxx GPUs, but older AMD cards will not.
They call it hurdles; I just call it AMD and Microsoft being AMD and Microsoft. This could actually benefit both the RX 6xxx series and the 7xxx series.
Will likely come after some time for the RX7000 series maybe a few months since it'll be driver driven but it should be a nice boost for 9000 and 7000series , devs must still implement correctly and not just overdue rt or textures now on purpose for we don't see gains for very slight quality increase .
Posted on Reply
#7
POLDOXII
SspacenaxeWill likely come after some time for the RX7000 series maybe a few months since it'll be driver driven but it should be a nice boost for 9000 and 7000series , devs must still implement correctly and not just overdue rt or textures now on purpose for we don't see gains for very slight quality increase .
From what I understand, it’s already available for AMD’s 7000 series as well as the 9000 series. The preview driver is this one: AMD Software: Agility SDK Developer Preview Edition 26.20.12.01
Posted on Reply
#8
jdrch
AGlezBvalue is the maximum under optimal/lab conditions
Yeah did you know your Ethernet, Wi-Fi, and 5G connections are "up to" also? Obviously YMMV depending on conditions, which testing can't completely account for.
Posted on Reply
#9
AGlezB
jdrchYeah did you know your Ethernet, Wi-Fi, and 5G connections are "up to" also? Obviously YMMV depending on conditions, which testing can't completely account for.
So? Have you ever heard any of those making a claim of 90% improvements in any variable? The truth of the matter is already in the post:
AleksandarKthough in real games, we can generally expect a few percentage points of improvement
Posted on Reply
#10
radosuaf
POLDOXIIFrom what I understand, it’s already available for AMD’s 7000 series as well as the 9000 series. The preview driver is this one: AMD Software: Agility SDK Developer Preview Edition 26.20.12.01
From the Release Notes:
  • AMD Radeon RX 9000 series graphics products will support:
    • DXR Updates: Clustered Geometry, Partitioned TLAS, and Indirect Builds
Posted on Reply
#11
Rulumi
PierreJGThe research I did was that, for example, the RTX 3060 will have day-one support; the same goes for the RX 9xxx GPUs, but older AMD cards will not.
They call it hurdles; I just call it AMD and Microsoft being AMD and Microsoft. This could actually benefit both the RX 6xxx series and the 7xxx series.
RDNA 2 is never going to get any of these or other newer features, their "Maintenance Mode" means that their driver code branch only gets minor changes for bugfixes and AMD actively tries to have no other work done to it unless strictly necessary, compilers, modules and any API updates included.
Posted on Reply
Add your own comment
Nov 18th, 2026 15:37 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts