I’ve always been in favor of smarter tech solving brute-force problems, especially when it comes to PC gaming. More efficient rendering, better upscaling, and clever compression — have all worked in the right direction for an industry that's constantly asking for more power. Nvidia's Neural Texture Compression is one such solution, and it aims to slash VRAM usage in games so drastically that it almost sounds magical.

While NTC promises to make games a lot lighter on memory, that doesn't mean it makes them easier to run across the board. Like most of Nvidia’s biggest breakthroughs, the real benefits of Neural Texture Compression climb upward instead of trickling down evenly, favoring the hardware that already has the most headroom.

Neural Texture Compression drastically reduces VRAM usage in games

It replaces traditional compression technology through AI

Nvidia's NTC, or Neural Texture Compression tech, is an entirely different take on the graphics rendering pipeline. Pretty much everything you see in a game — every tree, wall, face, or object — is essentially just a high-resolution image that is wrapped around a 3D object. As such, rendering them in real time when you play a game comes with a heavy GPU resource cost, and that's the biggest drain for your VRAM. NTC, on the other hand, is looking to replace that entirely.

Nvidia's Neural Texture Compression stores a compact neural representation instead of full high-resolution textures. Then, it reconstructs those textures in real time using a tiny AI model, and it uses way less VRAM. In essence, the load shifts from the GPU's CUDA cores over to the AI-backed Tensor cores instead. At GTC 2026, Nvidia unveiled the tech's real-life examples, showing texture data worth over 6GB of VRAM requiring just under 1GB using NTC instead of traditional Block Compression (BCn) methods.

Modern AAA games are already halfway through making 8GB VRAM cards obsolete. Of course, GPUs with 8GB VRAM still have their place in the global pre-owned market and even in esports games, but there's no denying that more new games are pushing against the 8GB VRAM ceiling every year. That's why everyone and their dog is fed up with Nvidia continuing to make and sell cards with such little VRAM. With NTC, however, future adoption might be a major reason that Nvidia continues to keep making and selling GPUs that ship with the same 8GB VRAM.

Less VRAM usage, but only for the newest GPUs

NTC won't magically make your aging GPU play games better

Credit: Nvidia

Neural Texture Compression is a new technology that is meant to replace older methods in the pipeline. This means that games like Hogwarts Legacy and Alan Wake 2, which are notorious for pushing VRAM limits in games, won't suddenly require way less VRAM. For NTC to become mainstream, developers will have to actively incorporate it into their games, which means that future games may come with a decreased VRAM cost. Current titles sitting in our libraries, however, will have the same resource cost as they always have.

There's also another problem, which is that NTC runs on Tensor cores instead of CUDA cores. It will require a GPU's AI and high-performance computing cores instead of the CUDA cores or Stream Processors. Now, something like an RTX 2070 Super may have 8GB VRAM and a mid-range CUDA core count, but its Tensor cores are nowhere near as strong or efficient as those in the newer AI-backed cards like the RTX 40 and RTX 50 series. Much like DLSS 4.5's image quality and temporal stability improvements, NTC also comes with a not-so-insignificant performance cost — the heavy VRAM compression requires more performance from the Tensor cores' matrix acceleration engines.

👁 XDA
Quiz
8 Questions · Test Your Knowledge

Nvidia's Neural Texture Compression technology
Trivia challenge

Think you know the truth about Nvidia's NTC — does it really need a top-tier GPU to shine?

NvidiaGraphicsAI TechHardwareGaming
01 / 8AI Tech

What does Nvidia's Neural Texture Compression (NTC) primarily use to compress and decompress textures?

Correct! NTC leverages neural networks to encode texture data into a compact representation that is decoded in real time using AI inference. This is a fundamentally different approach from classical block compression methods like BC7 or DXT.
Not quite. NTC stands apart from traditional methods by using neural networks rather than classical algorithms like DCT or entropy coding. The AI-driven approach is what gives it both its compression advantages and its hardware considerations.
02 / 8Hardware

Which hardware component within Nvidia GPUs is most critical for accelerating NTC's real-time decompression?

Correct! NTC's neural inference workload is heavily accelerated by Tensor Cores, which are designed specifically for matrix multiplication operations used in AI and machine learning tasks. Without Tensor Cores, performance drops significantly.
Not quite. While CUDA cores and TMUs play roles in standard rendering, it is the Tensor Cores that make NTC viable at real-time speeds. Tensor Cores handle the matrix math that neural network inference depends on.
03 / 8Nvidia

On which Nvidia GPU architecture were Tensor Cores first introduced, making it the earliest generation capable of accelerating NTC-style workloads?

Correct! Tensor Cores debuted in the Volta architecture, specifically with the V100 data center GPU. However, they became widely available to consumer audiences starting with the Turing (RTX 20 series) architecture.
Not quite. Tensor Cores first appeared in the Volta architecture with the V100. Pascal and Maxwell lacked them entirely, meaning those generations cannot efficiently accelerate NTC workloads the way newer GPUs can.
04 / 8Graphics

Compared to traditional block compression formats like BC7, what is one of the key advantages Nvidia claims for NTC?

Correct! Nvidia claims NTC can deliver comparable or superior visual quality to BC7 while achieving much smaller texture footprints on disk and in memory. This can be a major benefit for large open-world games with extensive texture assets.
Not quite. The headline advantage of NTC over BC7 is better quality-to-size ratio — not mipmap elimination or legacy hardware compatibility. The decompression step still happens at runtime; it just uses neural inference instead of fixed hardware logic.
05 / 8Gaming

Why might NTC perform poorly or be impractical on older or lower-end GPUs without dedicated AI acceleration?

Correct! Without Tensor Cores or equivalent AI acceleration hardware, the neural network inference required by NTC falls onto general-purpose CUDA shaders, which are far less efficient for this workload. This results in a significant frame rate penalty that makes NTC impractical on older hardware.
Not quite. The bottleneck is computational, not bandwidth or VRAM-related. When Tensor Cores are absent, the AI inference workload is handled by regular shaders, which dramatically reduces performance and makes NTC unsuitable for real-time use on those GPUs.
06 / 8AI Tech

Which of the following best describes what NTC stores on disk instead of traditional raw or block-compressed texture data?

Correct! NTC encodes textures as trained neural network weights rather than storing pixel data directly. At runtime, the GPU runs inference using those weights to reconstruct the texture, which is why the file sizes can be dramatically smaller than traditional formats.
Not quite. The elegant concept behind NTC is that the texture's content is baked into neural network weights. Instead of storing colors per texel, you store a compact model that can reproduce them — which is why it needs AI hardware to decode efficiently.
07 / 8Hardware

Nvidia's NTC was introduced as part of which broader SDK or developer toolset?

Correct! NTC is part of Nvidia's RTX Neural Shaders (RTXNS) initiative, which provides developers with tools to embed neural network inference directly into shader pipelines. It represents Nvidia's broader push to make AI a core part of the graphics rendering process.
Not quite. NTC falls under the RTX Neural Shaders umbrella, not DLSS or RTX Remix. While DLSS is also AI-powered, it focuses on upscaling rather than texture compression, and RTX Remix is a modding platform for classic games.
08 / 8Graphics

Which statement most accurately reflects Nvidia's NTC and GPU performance requirements?

Correct! This is the nuanced reality of NTC. It is not exclusively locked to flagship GPUs, but it performs best where Tensor Cores are present and powerful. On hardware without them, the fallback to shader-based inference makes the performance hit too steep for practical gaming use.
Not quite. NTC is not a hard requirement for a single GPU tier, nor is it equally efficient on all hardware. The truth is that Tensor Core availability and generation determine how practical NTC is — making it shine on modern RTX GPUs while struggling on older or lower-end ones.
Challenge Complete

Your Score

/ 8

Thanks for playing!

A decreased VRAM pressure doesn't automatically mean that the workload disappears, though. It just shifts the goalposts. Neural Texture Compression will trade memory bandwidth and capacity for real-time reconstruction on Tensor cores instead. That just means more compute cycles on the AI cores, resulting in potentially larger power draw, more heat, and potentially less headroom for everything else that a GPU would be trying to juggle while playing a modern game.

Only the new AI-centric GPUs will get the full benefit

NTC might hand Nvidia their best 8GB VRAM justification in years

The newest, most efficient Tensor cores are present, of course, in the newest GPUs from the brand, like the RTX 40 and 50 series. These cards make the most of their Tensor cores for AI-backed systems like upscaling using DLSS and frame generation as well. As such, it only makes perfect sense that if NTC were to become mainstream in game development going forward, it would work only on the cards with the most processing power and AI cores. This isn't just a guess, either. The Inference on Sample method of Neural Texture Compression, which drastically reduces VRAM usage, has already proven to work only on the most powerful Nvidia GPUs. For the rest of them, there's another NTC method called Inference on Load, which decreases texture size and PCIe traffic, but doesn't do anything to bring down the VRAM usage of a game.

That's why NTC could also prove to be Nvidia's best excuse to continue selling GPUs with 8GB VRAM in the future. Sure, modern titles are pushing back against cards with lower VRAM, but NTC could flip that on its head. I can almost hear CEO Jensen Huang saying they've made memory "smarter instead of larger" at the next conference as they reveal the next generation of RTX 60 series cards with 8GB VRAM.

This isn't something we haven't seen before

The best leaps in rendering tech are locked to newer hardware

Now, on the face of it, older GPUs like the RTX 20 series and RTX 30 series can use Nvidia's NTC tech when it comes to their games. After all, they do have the Tensor cores required for using DLSS. That's pretty much why even the latest iteration of Nvidia's upscaler, DLSS 4.5, is available for use on these older RTX cards. They may not be able to run modern AAA titles at the same visual fidelity as their newer counterparts, but they sure can use DLSS 4.5, even if it demands significantly more VRAM usage.

However, we know that much like frame generation, Nvidia's newest tech remains locked to their newer hardware, and even if the CUDA core count doesn't double every generation, the Tensor core numbers go up significantly as more of the rendering workload on cards is diverted to AI-backed processes. As such, while NTC could really benefit the industry by genuinely bringing down VRAM usage in GPUs with modest VRAM counts, it still would only be viable on Nvidia's latest cards that are built from the ground up with NTC support in mind. That's only par for the course now.

Gigabyte GeForce RTX 5070 Ti Eagle OC Ice SFF

Smarter memory, same old hierarchy

NTC's most meaningful gains show up where the hardware is already the strongest.

Neural Texture Compression is, without a doubt, one of the more interesting shifts we've seen in how games could be built and rendered going forward. It tackles a very real problem, and in isolation, it's hard not to appreciate the engineering behind it. But like most modern GPU innovations, it exists within a very specific hardware ecosystem.

In said ecosystem, the pattern remains the same, where the most meaningful gains show up where the hardware is already the strongest. Everyone else, meanwhile, gets a version of the feature that is technically there, but practically limited. NTC might change how VRAM is used, but it doesn't change who benefits the most from that change.