Abstract created by Good Solutions AI
In abstract:
- PCWorld stories that neural texture compression from Intel, Nvidia, and AMD might drastically scale back sport file sizes and VRAM utilization for PC avid gamers.
- Intel’s TSNC know-how compresses textures as much as 17X whereas Nvidia’s NTC reduces VRAM by over 85%, serving to older graphics playing cards run fashionable video games.
- Microsoft plans DirectX integration and Intel expects an alpha SDK launch this yr, doubtlessly fixing excessive RAM prices and storage points.
PC avid gamers face an ongoing drawback: extra highly effective video games demand extra highly effective sources, all within the service of video games that ship extra sensible experiences and graphics. However can avid gamers keep away from paying out outrageous sums of cash to maintain up?
Texture compression may be a solution, shrinking down the scale 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 could possibly be out there to new and current {hardware} within the coming months. All informed, it’s an thrilling potential answer 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 corporations 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 could possibly be successfully impressed utilizing its {hardware}, too.
3D graphics is basically a puppet present: Traditionally, every object is created by a framework of surfaces, after which the sport designers inform your PC the right way to “cowl” them with textures which might be individually lit and coloured. It’s this texture knowledge that may make up the majority of the sport’s measurement, 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 known as “maps.” And so they matter: A sport like Hogwarts: Legacy may require 58GB of knowledge; the “Excessive Definition Texture Pack” can require an extra 18.3GB. Loading textures out and in of reminiscence may trigger stuttering in video games, so decreasing the scale of these textures can also enhance the way in which the sport performs, too.
Microsoft, which is attempting to construct a DirectX API that can permit this to occur, has already mentioned that it plans to construct help for neural texture compression into DirectX. It’s assuming that builders will need to use what it calls each “small fashions” and “scene fashions” to permit next-generation scene rendering, which might embody 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 knowledge
Intel engineers confirmed its Texture Set Neural Compression working in two variants, compressing textures as much as 9X or over 17X versus uncompressed knowledge, relying upon the tactic used. Based on Intel graphics engineer Marissa Dubois, the textures could possibly be decompressed at varied factors: at set up, whereas the sport was loading, and even later. Like different compression methods, it makes use of similarities within the knowledge texture maps to attempt to scale back the info measurement.
Intel’s TSNC in motion.
YouTube / Intel
Intel’s presentation famous that there’s a little bit of what it calls “perceptual error: 6-7 p.c for the second 17X variant, or 5 p.c 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 technique, Intel mentioned.
For now, it is a demo, Intel engineers mentioned. An alpha software program growth package (SDK) is scheduled for later this yr, 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 technique to “enhance” the standard of video games — the place DLSS 5’s “enhancements” could be very a lot doubtful. Nvidia’s Neural Texture Compression is explicitly deterministic, which implies that it all the time reconstructs the identical texture that the developer designed.
Nvidia does use a small neural community to reconstruct the info, operating on its Tensor cores. The RTX Neural Texture Compression SDK is obtainable for builders to make use of at this time.
Nvidia confirmed off two demonstrations: neural texture compression and neural supplies. Nvidia demonstrated how NTC could possibly be used to compress a scene that beforehand required 6.5GB of VRAM to simply 970 MB with NTC. It’s not as clear what Nvidia is doing with neural supplies, however it seems the corporate is attempting to inform the GPU what the properties of a cloth truly are, after which compress these directions. The graphics card would then basically assemble the fabric, rushing up the method from between 1.4X to 7.7X, based on Nvidia.
Nvidia seems to be attempting to encode what materials act like in the actual world.
AMD doesn’t but provide an SDK to decrease reminiscence consumption in video games. Nonetheless, in 2024 it printed a paper on neural texture block compression, slicing texture sizes by 70 p.c.
All informed, neural compression isn’t one thing that may profit you but. However it seems to be just some months off…and let’s face it, it in all probability can’t get right here quickly sufficient.
