Direct3D Cooperative Vector is Evolving
DirectX continues to evolve to support the growing demands of AI and machine learning workloads. As part of this effort, we plan to introduce an expanded suite of linear algebra capabilities in a future Shader Model release, unlocking even more robust GPU acceleration for bedrock AI/ML computations within the graphics pipeline.
This work extends our experimental Cooperative Vector implementation by introducing accelerated matrix-matrix operations, crucial for scaling AI/ML workloads beyond what lightweight vector-matrix alone can handle. This will unlock new performance gains in compute shaders by enabling efficient, large-block tensor operations, making it ideal for demanding tasks like super resolution, LLM inference, and computer vision. Both matrix-matrix and vector-matrix capabilities will be unified under a single design, providing a consistent and scalable foundation for high-performance shader code for AI/ML scenarios.
Developer feedback has been central to this evolution, from early experimentation to shaping the unified design. If you have additional input, we’d love to hear it – contributions to the spec can be made here!
Development Notes
- While the unified design will not ship in Shader Model 6.9, it is planned for a future Shader Model release.
- Feature development has migrated to a new unified spec, linked here.
- To consolidate functionality and reduce fragmentation, the original Cooperative Vector spec will be deprecated. The feature itself will remain available as experimental until the replacement is released, at which point we will also provide a migration guide. In the meantime, we encourage developers to continue experimenting with Cooperative Vector and providing feedback to help inform the expanded feature.
SM6.9 Updates
We plan to release all other Shader Model 6.9 features for retail availability in Q1 2026. This includes:
Really gutted that this stuff can’t be done in a way to support AMD GPU’s Architecture atleast partially, esp with stuff like SER being no-op. Seems all or nothing 😕