Shader Model 6.10 has just entered preview alongside AgilitySDK 1.720, and it is a more notable DirectX 12 update than a routine API bump. Microsoft isn't promising higher FPS for gamers right away, but it is laying the groundwork for in-game shaders to handle matrices, split work across waves, and use GPU memory more flexibly. In short, this is preparation for neural rendering, where part of real-time graphics will be assisted by very small AI models running right inside the render pipeline. According to Microsoft's official announcement, the new preview package also ships with DXC 1.10.2605.2 and experimental drivers from AMD, Intel, and NVIDIA.
What does Shader Model 6.10 add, and why should PC gamers care?
The centerpiece of Shader Model 6.10 is LinAlg, a new linear algebra API for shaders. Put simply, it lets developers call vector and matrix operations closer to the hardware, instead of routing around hard-to-optimize code as before. For PC games, this change matters because many new graphics techniques need to process matrices very quickly on the GPU.
The preview also adds Group Wave Index and Variable Group Shared Memory. In practical terms, a shader now knows which processing group it belongs to within the GPU, and it can use a larger shared memory region if the hardware allows. Tasks that need many threads working together, which are the foundation of neural rendering, become less dependent on programming tricks that tend to behave differently across AMD, Intel, and NVIDIA GPUs.
Microsoft has also added new raytracing intrinsics such as TriangleObjectPositions and ClusterID, along with batched asynchronous command list calls to reduce serial bottlenecks in D3D12. The initial hardware support table shows NVIDIA with the broadest coverage, while AMD and Intel have only partial implementations that vary by GPU generation.
What stands out is that Shader Model 6.10 doesn't arrive as a standalone feature. It comes with Cooperative Vectors, LinAlg, and a more flexible memory management layer, so developers can bundle many small AI processing steps right inside the graphics pipeline. If game engine makers make good use of it, PC games over the next few years could both ease VRAM pressure and keep images more stable.
What do Cooperative Vectors and Neural Block Texture Compression signal for new GPUs?
If you need one easy-to-picture example, Neural Block Texture Compression is the clearest answer. In its post on DirectX Cooperative Vectors, Microsoft and NVIDIA describe a new approach: instead of storing textures in fixed compressed blocks and decompressing them the old way, a game can use a small neural network to reconstruct a texture at the moment a shader needs it. The intended benefit is lower VRAM pressure and smaller asset sizes while keeping image quality.
| Technology | Role | Benefit |
|---|---|---|
| LinAlg / Matrix API | Calls matrix operations more directly | Lays the groundwork for real-time AI shaders |
| Group Wave Index | Identifies a wave's position within its processing group | More stable work distribution across many GPUs |
| Variable Group Shared Memory | Expands shared memory according to GPU capability | Suits large workloads and AI texture compression |
| Neural Block Texture Compression | Uses a small neural network to represent textures | Reduces VRAM use and asset size |
At GDC 2025, Microsoft said Intel demonstrated Neural Block Texture Compression running about 10 times faster when taking advantage of Cooperative Vectors. That isn't a promise that games will soon load 10 times faster, but it shows this direction now has baseline figures behind it rather than being just an idea.
From an ecosystem standpoint, this move can also be seen as a piece behind the upscaling technologies advancing quickly on PC, from AMD FSR to NVIDIA's neural shading SDKs. The difference is that Shader Model 6.10 is not a feature players switch on in a game menu, but a lower-level standard layer that lets developers write a more durable code path for many GPUs. If the pace of drivers and tools stays steady, neural rendering on Windows will move closer to real-world use.
