Nvidia has a new tech called Neural Texture Compression. It cuts VRAM use a lot. In one demo, textures went from 6.5 GB down to 970 MB. That is about 7 times less memory. https://t.co/8F9ZJJjuKS
An animated infographic (GIF) from NVIDIA’s Developer Blog (Feb 6, 2025) that visually compares traditional block-compressed texture VRAM usage with RTX Neural Texture Compression (NTC). The GIF shows memory-usage comparisons for demo scenes (e.g., pilot helmet / Zorah) and illustrates the multi‑fold VRAM reductions NTC achieves, directly supporting the claim in the tweet about large VRAM savings (several× reduction).
Source: NVIDIA Developer Blog (developer.nvidia.com)
Research Brief
What our analysis found
NVIDIA's Neural Texture Compression (NTC) is generating significant buzz after demos at GTC 2026 showcased dramatic VRAM savings. In the most widely cited example, a "Tuscan Wheels" scene that consumed 6.5 GB of VRAM using conventional BCn block-compressed textures was reduced to just 970 MB with NTC — an approximately 6.7× reduction, or roughly 85–87% smaller. A separate "flight helmet" test pushed savings even further, with textures shrinking from 98 MB to approximately 11.37 MB, representing a 95.8% reduction in that specific scenario.
The technology works by using small neural network decoders (MLPs) to decompress physically-based rendering (PBR) texture bundles on-the-fly during rendering. NVIDIA claims NTC can achieve quality comparable to traditional compression (PSNR of ~40–50 dB) at significantly lower per-texel bitrates. The RTXNTC SDK, first released in beta in January 2026, is publicly available on GitHub and supports both "inference-on-sample" (real-time decompression) and "on-load" transcoding modes.
However, the technology is still in a preview stage with important caveats. The best performance requires RTX 4000 or 5000 series GPUs with tensor core and Cooperative Vector acceleration, which can speed inference 2–4× over standard paths. NVIDIA's own SDK documentation includes explicit "DO NOT SHIP" warnings for its DirectX 12 Cooperative Vector preview support, and early community testers have reported instability with preview drivers. Game engine integration also requires meaningful asset pipeline changes and shader modifications, meaning this is far from a drop-in solution for existing titles.
Fact Check
Evidence from both sides
Supporting Evidence
NVIDIA's own research confirms the technology and demo numbers
The official NVIDIA Research page for "Random-Access Neural Compression of Material Textures" describes the NTC method, its results, and its random-access capability for GPU texture use, directly supporting the claims in the tweet (source: research.nvidia.com).
The 6.5 GB to 970 MB figure comes from an official GTC 2026 demo
Multiple press outlets including VideoCardz and Tom's Hardware reported the Tuscan scene demo numbers on April 4, 2026, all citing NVIDIA's own presentation slides showing the ~6.7× VRAM reduction (source: videocardz.com).
The RTXNTC SDK is publicly available with reproducible examples
NVIDIA's GitHub repository for RTXNTC includes sample scenes, compression figures, and example applications that developers can run to verify the compression ratios, lending transparency to the claims (source: github.com/NVIDIA-RTX/RTXNTC).
Independent early adopters have observed large VRAM reductions
Third-party enthusiasts and YouTube channels experimenting with the SDK and preview drivers reported significant VRAM savings in their own test conditions, corroborating NVIDIA's claims in at least some scenarios (source: Tom's Hardware, Windows Central community coverage).
Contradicting Evidence
VRAM savings come at a computational cost
NTC requires running MLP neural network evaluations per-texel during rendering, which is computationally significant. The SDK warns that decompression-on-sample adds meaningful GPU overhead, and fallback implementations without tensor core or Cooperative Vector support are substantially slower (source: RTXNTC SDK README on GitHub).
The technology is still in preview and not production-ready
NVIDIA's SDK documentation explicitly includes a "DO NOT SHIP" warning for the DirectX 12 Cooperative Vector path, and the feature requires a pre-release DirectX Agility preview and a special NVIDIA developer preview driver (590.26), signaling this is not yet ready for shipping games (source: RTXNTC SDK README).
Early community testing revealed instability and variable results
Users running preview drivers and SDK builds reported crashes, driver issues, and inconsistent performance across different scenes and hardware configurations. Press coverage from Windows Central and others stressed that real-world results will vary significantly by scene, engine, and GPU (source: Windows Central).
Significant engine and asset pipeline work is required
NTC is not a drop-in solution for existing games. Assets must be prepared or transcoded to NTC format, and game engines must integrate NTC decompression shader paths, meaning widespread adoption will require substantial developer effort and cannot simply be patched into current titles (source: RTXNTC SDK documentation and community discussion).
Claims of zero quality loss are overstated
While NVIDIA targets BCn-comparable quality at lower bitrates, NTC involves lossy neural compression with configurable quality-bitrate tradeoffs, meaning some visual differences from uncompressed originals are inherent depending on the compression settings chosen.
Report an Issue
Found something wrong with this article? Let us know and we'll look into it.