Nvidia has demonstrated a technique called Neural Texture Compression that uses small neural networks to unpack game textures on the fly, reducing VRAM pressure and install sizes without visible loss in image quality.
The numbers from Nvidia's demos are striking. A Tuscan Villa scene running standard block compression required 6.5 GB of VRAM. NTC brought that down to 970 MB with the image described as identical.
A flight helmet model compressed from 272 MB uncompressed to 98 MB under block compression and then to 11.37 MB under NTC, roughly 24 times smaller than the original. Nvidia also claims the technique can deliver up to 4x higher resolution in the final render.
The company paired NTC with a related approach it calls Neural Materials, where neural networks handle some of the shader mathematics that would otherwise run through conventional rendering pipelines.
In its 1080p demo, Nvidia reported render times up to 7.7 times faster as a result. The technique runs on Tensor Cores, the matrix-acceleration hardware built into Nvidia's GPUs for AI workloads, meaning it does not compete with standard rasterisation performance for resources.
Related reading
- Nvidia says the open versus proprietary AI debate is the wrong argument
- From Vera Rubins, Blackwell, Nemotron, NemoClaw, to Isaac and Cosmos - here's a whistlestop guide to NVIDIA's…
- Meta and Arm to co-develop data centre chips built for AI workloads
The broader significance is that this is not solely an Nvidia story. Microsoft has standardised a related programming interface called Cooperative Vectors. Intel and AMD have both shown comparable demonstrations. The direction of travel across the industry is consistent, even if no game currently supports any of these approaches in a shipping title.
Implementation in actual games is expected soon, though no specific titles or release dates have been confirmed.
The recap
- Nvidia demos Neural Texture Compression reducing texture memory use.
- Demo cut VRAM from 6.5 GB to 970 MB.
- No game supports Cooperative Vectors or NTC yet.
