NVIDIA NTC Slashes VRAM Usage by 85% With AI Compression

A new AI-powered texture system could drastically reduce memory usage in games. NVIDIA has introduced Neural Texture Compression (NTC), a machine learning approach that changes how textures are stored and rendered.

AI replaces traditional texture storage

According to Wccftech, instead of storing full texture data like traditional formats, NTC compresses textures into compact latent representations. The GPU then reconstructs the final image in real time using a small neural network.

This process remains deterministic, meaning it produces consistent results every time rather than generating new variations.

How Neural Texture Compression works

NTC relies on a combination of compressed data and lightweight AI processing. The system stores a latent texture and uses a neural decoder to rebuild the final output during rendering.

Positional encoding helps preserve fine details, ensuring that reconstructed textures maintain visual accuracy even at high compression levels.

Training enables high-quality reconstruction

The neural network learns how to rebuild textures during a training phase. It optimizes both the latent data and model weights through repeated iterations, minimizing reconstruction errors.

This allows NTC to maintain quality while significantly reducing storage requirements compared to traditional compression methods like BCN.

Major gains in memory and efficiency

In NVIDIA’s demo, NTC reduced VRAM usage from around 6.5GB to just under 1GB. That translates to roughly an 85% reduction without sacrificing texture fidelity.

The system also supports multiple material channels such as normals and roughness in a single representation, avoiding the need to split data across multiple textures.

What this means for games

NTC could lead to smaller game downloads, faster updates, and better visual quality within the same hardware limits. Developers may also push higher-resolution assets without increasing VRAM demands.

The technology already works across NVIDIA, AMD, and Intel GPUs, making it broadly compatible despite being introduced by NVIDIA.

Still early, but promising

The SDK for Neural Texture Compression is currently available in beta, and no major titles use it yet. Adoption will depend on how easily developers can integrate it into existing pipelines.

While NTC focuses on texture efficiency, NVIDIA continues to evolve its upscaling tech. DLSS 5 remains under scrutiny, with reports suggesting a possible driver-level toggle.

At the same time, a DLSS 4.5 over-the-air update has surfaced, rumored to include up to 6x multi-frame generation.