Microsoft Overhauls DirectX 12 Ray Tracing

Microsoft unveiled significant advancements on October 28. They launched the official DirectX 12 Agility SDK version 1.620 alongside the experimental version 1.722 preview. As detailed in the Microsoft developer blog, these releases deliver a comprehensive suite of updates explicitly targeting DirectX Raytracing (DXR).

The experimental SDK introduces highly sophisticated features. These include Clustered Geometry, Partitioned Top-Level Acceleration Structures (PTLAS), and Indirect Acceleration Structure Building. The primary objective is to drastically reduce the enormous computational overhead associated with executing ray tracing in highly complex, dynamically shifting scenes. According to Microsoft, this represents the most monumental evolution of DXR since its initial inception in 2018.

The Mechanics of Clustered Geometry

Historically, ray tracing algorithms require the continuous construction and updating of acceleration structures. These mathematical frameworks allow the Graphics Processing Unit (GPU) to rapidly calculate the precise intersection between light rays and scene geometry. Typically, developers employ Bounding Volume Hierarchies (BVH) to manage this data. However, in environments brimming with intricate animations, streaming geometry, and rapidly transforming objects, repeatedly rebuilding these structures incurs an exceptionally high computational penalty.

To solve this, Microsoft introduces the Clustered Geometry feature. This technique elegantly subdivides vast geometric datasets into diminutive, reusable clusters. Specifically, each cluster encapsulates a maximum of 256 triangles and 256 vertices. Consequently, developers can seamlessly combine these modular clusters to construct much larger ray tracing acceleration structures.

This approach significantly diminishes the repetitive labor of processing redundant geometric data. Ultimately, it empowers creators to utilize exquisitely high-fidelity models for ray tracing with unprecedented efficiency. Developers can now confidently reuse the exact same high-resolution geometric data across both traditional rasterization and modern ray tracing pipelines. This eliminates the tedious necessity of fabricating simplified proxy models exclusively for ray tracing operations.

Advancements in Acceleration Structures

Another critical innovation is the Partitioned Top-Level Acceleration Structure (PTLAS). This feature specifically targets the update efficiency of sprawling, dynamic environments. Traditional Top-Level Acceleration Structures (TLAS) demand the holistic management of every object instance within a scene. When the overall scene scale is massive, even trivial changes to a few objects can generate exorbitant update costs.

PTLAS resolves this by permitting developers to logically segment the environment into discrete partitions. Consequently, they only need to update the specific partitions where actual changes occur. They can effortlessly preserve the existing structural integrity of the static, unchanged regions.

Furthermore, by integrating GPU-driven acceleration structure building capabilities, developers can authorize the GPU to directly participate in these construction and update procedures. This strategic offloading substantially alleviates the preparatory burden traditionally placed upon the Central Processing Unit (CPU).

Astonishing Performance Gains

According to benchmark data released by Microsoft utilizing the `D3D12RaytracingClusteredGeometry` example, the results are staggering. In a test configuration powered by an AMD Radeon RX 9070 graphics card, the implementation of Clustered Geometry propelled performance from a meager 1.2 frames per second (utilizing the conventional Bottom-Level Acceleration Structure) to an astonishing 79 frames per second. This represents an approximate 66-fold increase in performance.

Moreover, the time required to build the acceleration structure plummeted from 788 milliseconds to a mere 3.11 milliseconds. The total memory footprint dedicated to the acceleration structure also witnessed a dramatic reduction. It shrank from approximately 7.3 GiB down to just 774 MiB, equating to a near 90% decrease.

Hardware Support and Future Implementation

Microsoft emphasizes that these extraordinary figures are derived from highly specific developmental examples and testing configurations. They do not guarantee an identical magnitude of improvement across all commercial video games. The actual real-world efficacy will invariably depend upon the specific computational workload, the nuances of the game engine’s implementation, and the underlying GPU architecture.

Regarding hardware compatibility, Microsoft has announced an impressive initial support roster. This includes the entirety of the Nvidia GeForce RTX graphics card lineup and the AMD Radeon RX 9000 series. It also encompasses the discrete Intel Arc B-series and the integrated graphics found within select Intel Core Ultra processors. Microsoft confirmed that these novel DXR capabilities do not mandate entirely new hardware features beyond existing ray tracing proficiencies. Nevertheless, graphics card manufacturers must still provide the necessary driver support to unlock these enhancements.

Additional SDK 1.620 Enhancements

Beyond the monumental ray tracing updates, the official DirectX 12 Agility SDK 1.620 also refines Advanced Shader Delivery (ASD) technologies. Specifically, Partial Graphics Programs now permit developers to reuse pre-compiled shader components across multiple rendering pipeline configurations. This optimization holds the promise of significantly reducing both shader package sizes and compilation overhead. However, at this juncture, this specific optimization exclusively supports AMD Radeon RX 7000 and RX 9000 series graphics cards.

The 1.620 retail version also incorporates several other noteworthy additions. These include the DirectX Dump Files GPU crash diagnostic functionality, Fence Barriers Tier 1, and a batched asynchronous command list API. It also introduces the DirectX Compute Graph Compiler runtime interface, tailored specifically for machine learning workloads. Finally, Microsoft officially confirmed that Shader Model 6.10 remains entrenched in the preview phase. They plan to officially launch it alongside the highly anticipated Agility SDK 1.621 retail release, scheduled for early December.