Intel has been releasing more details about its Texture Set Neural Compression (TSNC), which uses AI to shrink texture data at install time or during runtime.
If it does what it says on the tin, it could be a big win for VRAM use in games.
All the big names in PC gaming hardware, plus Microsoft with DirectX, are pouring money into neural rendering tools that claim to make life easier for devs and less painful for gamers.
TSNC first showed up last year, but an updated version popped up at GDC, with Intel saying it plans to ship it later this year.
Like Nvidia’s Neural Texture Compression, TSNC targets the bulky footprint of traditional textures built around “block compression,” with AI compression taking a different route.
Intel claims TSNC can compress textures by up to 18X compared to the original format, which sounds wild until you remember how much modern games shovel into memory.
TSNC is planned as a standalone SDK that takes standard BC1-compressed textures, the industry workhorse format, then compiles them for decompression on modern GPUs and even CPUs.
One obvious use is cutting install and patch sizes. The neural network stores latent data to reconstruct textures, and it can run during installation, while a game loads, while streaming texture data or during per-pixel sampling at runtime.
That runtime path is where the VRAM savings are meant to land, while Intel says the tech can chase storage space, memory bandwidth and VRAM as separate goals.
In Intel’s current setup, TSNC supports two modes, Variant A and Variant B, with Variant B applying more compression.
Intel says Variant B hits up to an 18X reduction with only a seven per cent loss in visual fidelity.
The best-case path uses Intel GPUs with XMX, including Panther Lake mobile parts with integrated Intel Arc B-Series graphics, but there is a fallback for non-Intel GPUs and CPUs.
Chipzilla is clearly trying to make this look platform-friendly because it knows devs do not ship features tied to one badge unless there is cash on the table.
