‘Neural texture compression’ may save players in a RAM-starved world

PC players face an ongoing drawback: extra highly effective video games demand extra highly effective assets, all within the service of video games that ship extra real looking experiences and graphics. However can players keep away from paying out outrageous sums of cash to maintain up?

Texture compression may be a solution, shrinking down the dimensions of video games in addition to permitting them to suit into the restricted video reminiscence of older, cheaper playing cards. Each Nvidia and Intel are engaged on concepts, which might be out there to new and present {hardware} within the coming months. All instructed, it’s an thrilling potential resolution to the continuing shortage of RAM and video RAM, which is driving up costs (together with video playing cards) and holding again new graphics-card releases.

Each firms outlined their plans in current weeks: Intel introduced its Neural Texture Compression SDK this weekend, whereas Nvidia’s associated neural texture compression discuss at GTC 2026 confirmed how textures might be successfully impressed utilizing its {hardware}, too.

3D graphics is actually a puppet present: Traditionally, every object is created by a framework of surfaces, after which the sport designers inform your PC learn how to “cowl” them with textures which can be individually lit and coloured. It’s this texture information that may make up the majority of the sport’s dimension, since every object can have a number of “maps” utilized to it.

A sensible-looking “brick” may be coded to inform the sport which components of the brick are shadowed, that are tough or shiny, and the way these variations have an effect on the colour of the brick itself. These are referred to as “maps.” They usually matter: A sport like Hogwarts: Legacy may require 58GB of information; the “Excessive Definition Texture Pack” can require an extra 18.3GB. Loading textures out and in of reminiscence also can trigger stuttering in video games, so decreasing the dimensions of these textures can also enhance the best way the sport performs, too.

Microsoft, which is attempting to construct a DirectX API that may permit this to occur, has already stated that it plans to construct assist for neural texture compression into DirectX. It’s assuming that builders will wish to use what it calls each “small fashions” and “scene fashions” to permit next-generation scene rendering, which might embrace neural lighting and neural texture compression. In each, AI can be used to calculate how a scene ought to be drawn and shaded, versus truly performing all of those calculations itself.

Two methods of shrinking sport information

Intel engineers confirmed its Texture Set Neural Compression working in two variants, compressing textures as much as 9X or over 17X versus uncompressed information, relying upon the strategy used. In line with Intel graphics engineer Marissa Dubois, the textures might be decompressed at numerous factors: at set up, whereas the sport was loading, and even later. Like different compression strategies, it makes use of similarities within the information texture maps to attempt to scale back the info dimension.

Intel’s presentation famous that there’s a little bit of what it calls “perceptual error: 6-7 % for the second 17X variant, or 5 % with the primary. Each can both use the XMX cores contained in the Intel Arc GPUs, or “fall again” to a extra generic implementation that can be utilized on different CPUs and even GPUs. XMX inference on Panther Lake is about 3.4 instances sooner than the fallback methodology, Intel stated.

For now, this can be a demo, Intel engineers stated. An alpha software program improvement package (SDK) is scheduled for later this 12 months, adopted by a beta after which an eventual launch.

Nvidia has already unveiled DLSS 5, the controversial graphics enhancement which provides generative AI as a option to “enhance” the standard of video games — the place DLSS 5’s “enhancements” may be very a lot unsure. Nvidia’s Neural Texture Compression is explicitly deterministic, which signifies that it at all times reconstructs the identical texture that the developer designed.

Nvidia does use a small neural community to reconstruct the info, working on its Tensor cores. The RTX Neural Texture Compression SDK is out there for builders to make use of as we speak.

Nvidia confirmed off two demonstrations: neural texture compression and neural supplies. Nvidia demonstrated how NTC might be used to compress a scene that beforehand required 6.5GB of VRAM to only 970 MB with NTC. It’s not as clear what Nvidia is doing with neural supplies, nevertheless it seems the corporate is attempting to inform the GPU what the properties of a fabric truly are, after which compress these directions. The graphics card would then primarily assemble the fabric, rushing up the method from between 1.4X to 7.7X, in line with Nvidia.

AMD doesn’t but provide an SDK to decrease reminiscence consumption in video games. Nevertheless, in 2024 it revealed a paper on neural texture block compression, slicing texture sizes by 70 %.

All instructed, neural compression isn’t one thing that may profit you but. Nevertheless it seems to be just some months off…and let’s face it, it most likely can’t get right here quickly sufficient.