Argomenti trattati
The modern PC game balances visual fidelity against the practical limits of storage and memory, and textures are often the largest contributors to a game’s footprint. Developers apply multiple image maps—color, roughness, normal maps and more—to give surfaces realistic appearance, and those maps consume gigabytes of space when stored at high resolution. Reducing that burden without visibly degrading image quality has been a long-standing challenge, and now several major vendors are pursuing a family of techniques known broadly as neural texture compression to address it. These AI-based methods offer a path to much smaller downloads and less frequent texture streaming stalls that cause in-game stutter.
Beyond raw file size, texture data also competes for limited VRAM on graphics cards, forcing engines to swap assets in and out and sometimes preventing older GPUs from running modern titles at reasonable settings. Game upgrades such as optional high-definition packs can add tens of gigabytes to an install, demonstrating how quickly texture sets balloon overall storage needs. Hardware and middleware vendors argue that neural compression can keep image fidelity close to original while compressing content, and that integrating these codecs into graphics APIs and engines will let both new and legacy systems benefit without wholesale hardware replacement.
Why texture compression matters
In many game projects the majority of shipped bytes are images: diffuse color maps, normal maps, ambient occlusion and specular maps together make up massive texture atlases. When engines stream those assets at runtime they can cause frame hiccups if the GPU memory cache misses and the system must load large files from disk. Compressing textures more aggressively reduces both on-disk size and active memory pressure, improving load times and reducing stutter. The promise of neural texture compression lies in using small neural decoders to recreate high-quality imagery from compact encodings; these decoders can run on specialized cores or CPU fallbacks, reconstructing the textures deterministically or with controlled perceptual differences as needed.
Industry approaches to neural compression
Intel’s TSNC and deployment plan
Intel’s prototype, called Texture Set Neural Compression (TSNC), offers multiple operating points: one approach compresses texture data by roughly nine times versus uncompressed assets, while an alternative mode reaches in excess of seventeen times compression with a measurable trade-off. Intel engineers reported a small perceptual error—for their highest-compression mode they observed roughly 6–7 percent perceptual discrepancy compared with originals, and about 5 percent for the other mode—metrics used to balance size against visible quality. TSNC can run on Intel’s XMX inference hardware inside Arc GPUs, or fall back to general-purpose CPU/GPU code; Intel noted that XMX on its Panther Lake silicon is about 3.4x faster than the generic fallback. The company has indicated an alpha SDK will be available later this year, with subsequent beta and final releases planned.
Nvidia’s deterministic NTC and neural materials
Nvidia has already exposed developers to its own toolkit: the RTX Neural Texture Compression (NTC) SDK is available now and uses small networks executed on Tensor cores to reconstruct textures. Unlike generative systems that may produce variable outputs, Nvidia emphasizes a deterministic reconstruction so that developers can expect consistent results. In demos the company compressed a scene that previously consumed 6.5GB of VRAM down to roughly 970MB using NTC, demonstrating significant runtime memory savings. Nvidia is also exploring neural materials, a concept where material properties are encoded and then synthesized by the GPU at render time, accelerating material setup by factors reported between about 1.4x and 7.7x in their tests.
Practical implications for developers and players
Microsoft is preparing to fold this work into the graphics stack, announcing plans to add support for neural texture compression within DirectX. The API-level support envisions two model classes—small models and scene models—to give developers flexibility in how and where neural techniques are applied, ranging from localized texture decoding to broader scene-level inference such as neural lighting. Meanwhile, AMD has published research—most notably a 2026 paper showing about 70 percent reductions using neural-inspired block compression—though it has not released a public SDK for memory reduction in shipped games. If these vendor SDKs and API hooks mature, players could see smaller downloads, lower VRAM requirements on older GPUs, and fewer streaming hitches without a dramatic drop in visual quality.