Nvidia’s RTX Neural Texture Compression Slashes VRAM Use by Over 80% in Benchmarks

Nvidia’s RTX Neural Texture Compression (NTC) isn’t just a lab curiosity—it’s already delivering measurable wins on real hardware. In the Tom’s Hardware benchmark, a suite of games and synthetic scenes ran on everything from a desktop RTX 4090 to an RTX‑compatible ultrabook, and the VRAM footprint shrank by more than 80 % across the board. The test harness toggled the three DirectX 12 modes—Inference on Load, Inference on Sample, and Inference on Feedback—while the Vulkan runs were limited to the first two, per Nvidia’s own documentation. Even the most texture‑heavy titles saw their memory consumption dip from several gigabytes to under a gigabyte, a shift that could free up space for higher‑resolution assets or larger scene graphs without touching the GPU’s core clock.

The magic lies in how NTC repackages a texture. During compression, the original image is broken down into a compact set of neural‑network weights and latent features. At runtime, the GPU’s Tensor Cores execute a tiny multi‑layer perceptron (MLP) to reconstruct each texel on demand—a process Nvidia calls “Inference on Sample.” Because the MLP is deterministic rather than generative, the output matches the source pixel‑perfectly, sidestepping the visual artifacts that have plagued older lossy codecs. Tom’s Hardware notes that the latency added by this per‑texel inference is negligible on RTX‑50‑series silicon, thanks to the Cooperative Vectors engine that pipelines the MLP calls alongside traditional shading work.

Alexey Panteleev, a Distinguished DevTech Engineer at Nvidia and one of the architects behind NTC, explained that the approach was designed to “make part of the graphics pipeline trainable,” allowing developers to replace hand‑crafted shader logic with small, learned models. In practice, this means a game studio can train a network on a specific material library, embed the resulting weights into the texture bundle, and let the GPU handle decompression automatically. The Tom’s Hardware team verified this workflow by swapping out conventional DXT5 textures for their NTC equivalents and watching the VRAM meter tumble without any perceptible drop in frame rate.

Beyond raw memory savings, the compression scheme opens doors for new rendering tricks. Because the latent features are stored alongside the neural weights, developers can query them for additional data—such as material roughness or anisotropy—without pulling extra texture maps. This “dual‑use” capability could streamline pipelines that currently juggle dozens of texture slots, a point highlighted in the benchmark’s analysis of a high‑fidelity indoor scene where the number of active texture units dropped by roughly a third.

The verdict from Tom’s Hardware is clear: RTX Neural Texture Compression is ready for prime time, at least on Nvidia’s latest GPUs. With an 80 %+ reduction in VRAM usage and virtually no performance penalty, the technology promises to extend the life of existing hardware while giving developers a fresh tool for tackling ever‑growing asset budgets. As the industry pushes toward higher‑resolution displays and more complex virtual worlds, the ability to squeeze more visual fidelity into the same memory envelope could become a decisive advantage—one that Nvidia is already capitalizing on with its RTX 50‑series launch.