Microsoft previews Shader Model 6.10 with a matrix math API, making neural rendering a standard DirectX feature

Microsoft has released a preview of Shader Model 6.10 in the new AgilitySDK 1.720-preview build. The update brings several changes, including tweaks to shared shader memory management and ray tracing, but the headline addition is a new Matrix feature.

According to the developer blog, Shader Model 6.10 introduces a streamlined matrix algebra API through the linalg::Matrix class, exposing a full set of matrix operations across GPUs from AMD, Intel, and NVIDIA. Modern GPUs already include dedicated hardware for AI workloads, such as NVIDIA's Tensor cores, Intel's XMX units, and similar accelerators on AMD hardware. However, DirectX support for these capabilities has been limited until now.

With this update, developers can access those matrix units through a unified API, rather than relying on vendor-specific implementations. This allows neural rendering operations to be executed across multiple GPUs with a single programming effort, rather than as a proprietary feature that developers must implement separately for each graphics card family. Microsoft is essentially trying to make matrix math a standard part of the DirectX API, making neural rendering a core DirectX feature rather than something bolted on by individual GPU makers.

3

Matrix math is the computational paradigm used by modern AI systems, especially transformer models and most LLMs. NVIDIA GPUs have had dedicated matrix-math hardware since the RTX 20 series, when Tensor cores first appeared. AMD was slower to follow, only adding designated matrix math hardware with RDNA 4.

It's worth noting that all of this applies primarily to GPUs that support the new Microsoft API. Older AMD GPUs without matrix-math hardware are unlikely to qualify. Beyond the matrix addition, the Agility SDK 1.720-preview includes Variable Group Shared Memory. This increases shared memory limits from 32KB to whatever the hardware supports, allowing shaders to use the full GPU capacity. Some ray tracing intrinsics, specifically TriangleObjectPositions and ClusterID, have also been updated alongside batched asynchronous command list APIs.

3

That said, this should give us an idea of neural rendering's potential and the features that could become standard across future DirectX GPUs. Machine learning may soon stop being a post-processing trick tacked on at the end of a frame and instead tie directly into how games are rendered. That might set off DLSS 5 and AI slop alarm bells for some, but Shader Model 6.10 should still be a net positive because, in theory at least, it makes neural rendering less of an Nvidia-exclusive technology. Shader Model 6.10 is currently in preview and requires specific developer drivers from NVIDIA, though AMD and Intel's public drivers already have support.