Nvidia reduces 6.5 GB VRAM to 970 MB with Neural Texture Compression

Nvidia, GTC 2026 highlighted within the scope of Neural Texture Compressionin short NTCchanges the method of storing textures in game graphics. Instead of storing textures as full resolution pixel data, the company breaks them into smaller pieces and extracts the needed data at run time. GPU It rebuilds it with a small neural network on it. In the examples shared by Nvidia, in a scene VRAM usage 6.5GB from level 970MB goes down to the level. These data directly concern the memory consumption of high-resolution texture packages.

However, Nvidia is not using a classically productive model here. According to the information provided by the company, the system gives the same output to certain inputs every time and therefore works deterministically. This approach makes it easier for game engines to produce predictable results. On the other hand, the technology stands out from traditional compression methods because it reconstructs the texture data instead of moving it as it is.

Neural Texture Compressiondoesn’t just focus on reducing file size. In Nvidia’s technical explanation, the system belongs to a material. PBR It takes the textures together and transforms them into forms of representation to be resolved in “shaders”. The small neural network then reconstructs the needed “texel” information in a real-time workflow. Thus, texture data takes up less space on the storage side, and the required section is processed directly at runtime.

In addition, Nvidia states in the official SDK documents that this technology can compress thousands of textures in a short time. Compared to traditional block-based texture compression, depending on the scenario used, the company to 7 floors up to, on some official pages up to 8 times He talks about efficiency. This difference varies depending on the sample and measurement method used. The example of the scene shown by the company around GTC is directly 6.5GB And 970MB highlights its values.

Microsoft is preparing the necessary infrastructure

However, this technology is not included in any commercial games today. Nvidia, NTC solution RTX Kit It offers it to developers and expands its software tools within this framework. On the other hand, Microsoft also makes these types of workloads possible. Cooperative Vectors It first announced its approach in 2025. The company then installs the relevant components February 26, 2026 on Shader Model 6.9 And Agility SDK 1.619 Ported to retail version via .

Thus, the texture compression technique described by Nvidia is not just a technology demo. This foundation prepared by DirectX opens the door for small neural networks to work in shader stages. at this point neural rendering Under the title, not only the texture side but also different graphic workflows come to the fore. NTC provides a concrete example of direct memory usage within this topic.

Of course, this development is especially 8GB VRAM It is of particular importance for graphics cards in its class. In recent years, high-resolution textures, open-world games and large material packages have placed tougher demands on memory. Therefore, storing texture data in a smaller space allows developers to use more content within the same memory budget. In addition, texture transfer and memory printing within the game are directly connected to this area.

Since there is no direct application in games on the market yet, the technology is currently progressing with preparations on the part of the developer. Nvidia, RTX Kit, Neural Shaders And Neural Texture Compression explains its titles within the same framework. If Microsoft Cooperative Vectors And Shader Model 6.9 expands the software basis of this approach. Like this Neural Texture CompressionIt is among the texture compression options that may enter game engines in the future.