DirectX 12 Agility SDK 1.722: 66× Ray Tracing FPS Boost, 90% VRAM Cut

DirectX 12 Agility SDK 1.722 Preview: Clustered Geometry and PTLAS

Microsoft released DirectX 12 Agility SDK 1.722 preview, the most significant update to DirectX Raytracing since its 2018 debut. The SDK adds two core features:

Clustered Geometry splits 3D models into clusters of no more than 256 triangles and 256 vertices. Developers no longer need to rebuild the entire acceleration structure when geometry changes; they can reuse or rebuild only the modified clusters.

PTLAS (Partitioned Top-Level Acceleration Structure) extends the same idea to the scene level. The GPU updates only the regions that actually change, which is especially effective for open worlds and scenes dense with dynamic objects.

Performance Demonstration on AMD Radeon RX 9070

Microsoft benchmarked a test scene using an AMD Radeon RX 9070:

Traditional acceleration structure rebuild: 1.2 FPS .

With Clustered Geometry: 79 FPS — a 66× improvement .

GPU acceleration structure build time dropped from 788 ms to 3.11 ms .

Total acceleration structure VRAM fell from 7519 MiB to 774 MiB (~ 90% reduction ).

Per-frame ray tracing render time decreased from 72.6 ms to 9.27 ms .

Dynamic object acceleration structure VRAM plunged from 6794 MiB to 244 MiB ( 96.4% reduction ).

The trade-off is a modest increase in temporary VRAM required during the build process.

Hardware Support

NVIDIA : All GeForce RTX GPUs, including the original RTX 20 series.

AMD : Radeon RX 9000 series.

Intel : Arc B-series discrete GPUs and Core Ultra 200V/300 integrated graphics.

Qualcomm : Plans support for Snapdragon X2 in H1 2027.

Microsoft emphasizes that these features do not require new ray tracing hardware — only driver updates.

Ecosystem Adoption

The update standardizes NVIDIA's earlier RTX Mega Geometry technology into DirectX. Gears of War: E-Day has already adopted a similar approach. Microsoft and NVIDIA also released a path-traced forest demo covering 25 million square meters with over 60 million plants.

Agility SDK 1.620 Stable and Shader Model 6.10

The simultaneously released Agility SDK 1.620 stable adds Partial Graphics Programs for advanced shader delivery, currently limited to Radeon RX 7000 and RX 9000 series. Shader Model 6.10 remains in preview; a stable release is expected with SDK 1.621 in early December.

Microsoft cautions that the quoted performance figures come from specific development samples and configurations; real-world gains will vary by workload, implementation, and GPU architecture.

Why Setting Task Manager to High Priority Doesn't Guarantee Responsiveness

Addressing a common user question, Microsoft senior engineer Raymond Chen explains why raising Task Manager's priority to "Realtime" or "High" often fails to revive a frozen system.

Core Argument: Shared Resources Are the Real Bottleneck

In Windows scheduling, even a thread with a high priority cannot bypass its dependence on shared resources. Chen uses a supermarket analogy: you may have priority checkout access, but if the only cashier is busy scanning a cart full of items, you still wait. In the OS, those "cashiers" are mutexes, file handles, and network requests.

Priority Inversion Protection Has Limits

Windows does implement priority inversion protection — temporarily boosting the priority of the thread holding a needed resource so it releases the resource faster. However, if that thread is executing an uninterruptible long-running operation, the high-priority Task Manager remains blocked for an extended period.

UI Programs Are Especially Vulnerable

Both the taskbar and Task Manager are UI applications. They must ultimately enter the Windows unified UI subsystem for scheduling, competing with every other UI thread in the system. If the UI subsystem deadlocks because of a runaway program, Task Manager cannot escape the contention.

Fundamental Limitation

Attempting to repair a crashed system from within using a tool that shares core resources with the offending process is inherently difficult. Granting that tool higher priority does not help when the shared resource itself is the bottleneck. Windows currently prioritizes the foreground (focused) application, which Chen describes as the practical limit of scheduling efficiency.

}

Code example

来源丨硬件世界
作者
丨黑白