NVIDIA’s Neural Texture Compression Cuts VRAM Usage by 85% — From 6.5 GB Down to 970 MB
NVIDIA’s Neural Texture Compression Cuts VRAM Usage by 85% — From 6.5 GB Down to 970 MB
- 60% of MD5 Password Hashes Can Be Cracked in Under an Hour with a Single GPU
- Dirty Frag: Root Access on Every Major Linux Distribution — No Patch, No Warning
- Ubuntu 26.04 LTS (Resolute Raccoon): The Most Ambitious Ubuntu LTS in a Decade
- Proton Mail: Data Transferred to FBI Again!
- How Close Are Quantum Computers to Breaking RSA-2048?
- How to Prevent Ransomware Infection Risks?
- What is the best alternative to Microsoft Office?
NVIDIA’s Neural Texture Compression Cuts VRAM Usage by 85% — From 6.5 GB Down to 970 MB
At GTC 2026, NVIDIA detailed a technology three years in the making that uses small on-GPU neural networks to decompress textures in real time — slashing video memory requirements without sacrificing a pixel of visual quality.
During a session titled “Introduction to Neural Rendering” at GTC 2026, NVIDIA’s Senior DevTech Engineer Alexey Bekin walked developers through the company’s Neural Texture Compression (NTC) — a machine-learning approach to texture storage that has been in development for nearly three years and available via SDK since early 2026, but has yet to see adoption in a shipped game title.
The timing of the renewed spotlight appears deliberate. With AAA game textures routinely exhausting 8 GB of VRAM on mid-range graphics cards, NTC offers a compelling answer — and NVIDIA clearly wants developers to start paying attention.
How It Works
Traditional texture compression in games relies on fixed block-based formats, the BCn series, which divide images into rigid 4×4 pixel blocks. These formats are fast and hardware-supported, but they impose a quality floor: compress too aggressively and artifacts appear; compress less and VRAM fills up.
NTC takes a fundamentally different approach. Rather than storing every texel directly, it encodes textures into a compact set of learned latent features that capture their essential visual content. At runtime, a small neural network running on the GPU reconstructs texel values on demand — decoding directly in the shaders that would normally sample the texture. The critical enabling hardware is the matrix-acceleration engine already built into modern GPUs: Tensor Cores on NVIDIA cards, with equivalent units on AMD and Intel silicon.
NTC does not consume standard rendering resources. The neural decoding pipeline operates alongside traditional rendering work — not instead of it.
The Tuscan Wheels Demo
NVIDIA’s headline demonstration used a test scene called “Tuscan Wheels” — a detailed interior environment heavy with high-resolution surface textures. Under conventional BCn compression, the scene’s textures consumed 6.5 GB of VRAM. With NTC enabled and image quality visually unchanged, that figure dropped to 970 MB: an 85% reduction.
The compression benefit runs in both directions. Developers can use the freed VRAM headroom to reduce install size and lower hardware requirements — or they can spend it on finer detail, pushing texture resolution 4× higher while staying within the same memory budget.
Neural Materials: Compression Meets Rendering
Alongside NTC, NVIDIA detailed a related technology called Neural Materials. Where NTC targets texture storage, Neural Materials replaces the traditional BRDF (Bidirectional Reflectance Distribution Function) lighting calculations that determine how surfaces respond to light. A material set previously requiring 19 texture channels is compressed to just 8, and the lighter neural evaluation path yields rendering speed improvements between 1.4× and 7.7× at 1080p resolution — depending on scene complexity.
Both technologies operate on the same core principle: encode high-dimensional data into a compact learned representation, then decode it efficiently at runtime using dedicated matrix hardware.
Not Generative AI — By Design
NVIDIA was explicit in addressing potential concerns: NTC is not a generative AI system. The small neural networks involved are trained offline during the game’s development phase, on fixed texture assets specific to that game. Nothing is generated or hallucinated at runtime. The decoder simply reconstructs texture data from learned features — a deterministic, lossless-quality process that eliminates the risk of AI-introduced visual artifacts.
Cross-Platform and Cross-Vendor
One of the more significant aspects of NTC is its hardware agnosticism. While the technology was developed by NVIDIA and relies on matrix-acceleration engines, it is explicitly supported on AMD and Intel graphics hardware as well. The SDK is already available to developers on GitHub.
Hardware leaker Kepler_L2 has suggested that Sony’s PlayStation 6 console may also adopt NTC — an approach that would allow Sony to pair a 1 TB SSD with much smaller per-game install sizes without sacrificing visual fidelity.
Key Takeaways
- NTC cuts VRAM usage by up to 85% — a 6.5 GB scene compressed to 970 MB in NVIDIA’s demo
- Real-time decoding runs on Tensor Cores (NVIDIA) or equivalent matrix units (AMD, Intel)
- Developers can use the savings to lower requirements or raise texture quality by 4×
- Neural Materials delivers up to 7.7× faster 1080p rendering by compressing material channel data
- NTC is not generative AI — networks are trained at dev time on fixed assets, no hallucinations
- SDK in beta on GitHub; Unreal Engine 5 integration expected by late 2026
- No shipping game titles yet use the technology, but large-scale adoption is described as imminent
Outlook
NVIDIA’s NTC SDK is currently in beta on GitHub, and the company has signalled that expanded engine support — including Unreal Engine 5 — is expected before the end of 2026. Rumours point to several upcoming titles being first to ship with full NTC integration, which would give GPU owners with 8 GB or 12 GB of VRAM meaningful relief without requiring a hardware upgrade.
Whether the technology delivers on its considerable promise in real-world game conditions — with dynamic scenes, moving objects, and complex lighting — remains to be seen in production. But on the evidence of NVIDIA’s GTC demonstrations, Neural Texture Compression represents one of the more substantive advances in GPU memory efficiency in years, and the fact that it is already cross-vendor and heading toward PS6 suggests the industry is paying close attention.
