Microsoft recently revealed that Cooperative Vectors will be available in DirectX. The company will demonstrate this feature at the 2025 Game Developer Conference’s Advanced Graphics Summit, which will be held in San Francisco from March 17–21.
With cooperative vectors, matrices with vectors of any size can be multiplied on any shader step. As a result, they can be utilized for real-time inference, and more especially, per-pixel inference, to speed up the execution of neural rendering approaches on hardware with AI acceleration, in addition to training AI models, such as through the use of compute shaders. Developers can now fully utilize the capabilities of the XMX units on contemporary Intel GPUs with this new functionality. Discrete GPUs like Intel Arc (A and B series) and integrated GPUs in Intel Core Ultra Processors (Series 1 and 2) will support cooperative vectors.
In order to showcase the advantages of Cooperative Vectors on Intel devices, Intel is thrilled to be taking the stage alongside Microsoft. Intel’s Graphics Research team’s Anis Benyoub will describe how this new DirectX feature allowed for a 10x improvement in Neural Block Texture Compression’s inference performance.
Why Neural Block Texture Compression?
These days, most gaming textures have a resolution of 4K and are made up of numerous channels that encode surface characteristics like ambient occlusion, metalness, roughness, normal, albedo, and so on. For instance, the T-Rex asset in Figure 1 requires eight 4096 x 4096 PBR textures, each with 13 distinct channels. This means that each texture requires more than 200 million texture values to be stored in video memory and that’s just for one asset!
Such volumes of texture data can overwhelm storage capabilities on current-generation consoles, even with the aid of hardware texture compression (such Block Compression or BC). In order to save more than 100GB of disk space, many games, including Call of Duty, must constantly stream texture content from the internet (see: “Extending In-Game Textures using CDNs for Call of Duty: MWII” by Chris Fowler, GDC 23). Therefore, increasing texture compression is a significant problem, and it turns out that integrating AI with Cooperative Vectors is a fantastic approach to do this!
Neural Block Texture Compression
The fact that standard texture compression is limited to textures with up to four channels (RGBA) is one of its main drawbacks. Neural Block Texture Compression, on the other hand, uses and benefits from an arbitrary number of channels. As seen in Figure 2 below, all 13 channels are correlated, which means that they share structures and patterns across channels, which makes sense for, say, that T-Rex asset. Up to five times the compression ratio of a traditional BC compressor can be achieved by capturing and encoding these correlations in a neural network.
Cooperative Vectors Acceleration
In neural compression technique, texture channels are compressed to a huge matrix that it use as a Multi-Layer Perceptron, which is quite similar to the Ubisoft method (see “Real-Time Weinreich et al., “Neuronal Materials using Block-compressed Features.” As a result, decompression turns into an inference made up of a few vector/matrix multiplication operations. Intel may leverage AI hardware directly and increase computation speed by up to ten times on Intel Arc (B series) GPUs with the new DirectX Cooperative Vectors API.
DirectX Neural Rendering: Cooperative Vector Support Coming Soon
A Novel Approach to 3D Graphics Programming: Neural Rendering Neural rendering technology is a major advancement in the rapidly evolving field of 3D graphics. The collection of methods that use AI/ML to significantly alter conventional graphics pipelines is known as neural rendering. These novel techniques aim to expand the realm of real-time graphics’ potential. Cooperative vectors are at the heart of DirectX’s commitment to enabling neural rendering approaches across platforms.
It is thrilled to announce that intend to enable cooperative vectors in DirectX, which will enable the next generation of neural rendering approaches across platforms.
What are Cooperative Vectors, and why do they matter?
The performance of neural rendering techniques will be immediately enhanced by cooperative vector support, which will speed up AI workloads for real-time rendering. It will accomplish this by allowing matrices to be multiplied by vectors of arbitrary sizes, which optimizes the matrix-vector operations needed in huge quantities for AI inferencing, training, and fine-tuning.
A modest neural network may operate in a pixel shader without using up the entire GPU with cooperative vectors, which also allow AI activities to perform in separate shader stages. Developers will be able to easily incorporate neural graphics techniques into DirectX applications and expand access to AI-accelerator hardware across many platforms with cooperative vectors. Intel goal is to give game developers the state-of-the-art resources they require in order to produce the upcoming generation of immersive experiences.
What’s Next For Neural Rendering?
In order to provide cross-vendor support for cooperative vectors into the DirectX ecosystem, the HLSL team is collaborating with AMD, Intel, NVIDIA, and Qualcomm. For more information about cooperative vectors and its impending Preview release, stay tuned!
The power of Tensor Cores with neural shading in NVIDIA’s upcoming RTX 50-series hardware will be unlocked by cooperative vectors. Neural shaders can be used to produce photo-realistic game characters, better arrange geometry for better path tracing performance, and depict game elements with AI.
In conclusion
One interesting approach to lowering texture memory requirements is to combine neural textures. Neural textures can now use hardware AI acceleration on contemporary GPUs to achieve realistic real-time performance with new support for DirectX Cooperative Vectors. In order to give gamers a visually stunning and high-performing gaming experience, Intel is happy to implement this capability in Intel Xe GPU architectures.
