NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB - TechPowerUp

요약

NVIDIA has released more details about its Neural Texture Compression (NTC) technology, which significantly reduces GPU VRAM usage by up to seven times. In a technology demo presented during one of the GTC 2026 sessions, NVIDIA revealed that its Neural Textur…

본문

Sunday, April 5th 2026

NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB by AleksandarK Yesterday, 02:11 Discuss (63 Comments) NVIDIA has released more details about its Neural Texture Compression (NTC) technology, which significantly reduces GPU VRAM usage by up to seven times. In a technology demo presented during one of the GTC 2026 sessions, NVIDIA revealed that its Neural Texture Compression can reduce VRAM usage from 6.5 GB to just 970 MB in certain scenes. This was demonstrated in a video comparing a Tuscan Villa and its interior. With virtually no difference in texture appearance, Neural Texture Compression represents a major advancement in maintaining graphics fidelity while freeing up GPU memory for more game content. For instance, in both the exterior of the Tuscan Villa and the interior demo showcasing detailed tableware, NTC technology provides users with high-quality textures while maintaining excellent material quality.

NTC technology is an AI-driven texture output that allows games to feature high-quality complex materials without a performance penalty. Games can benefit from the substantial VRAM reduction that NTC offers while maintaining the same texture quality. Traditionally, games use block-compressed formats like BCn, such as BC5, BC6, or BC7, which are commonly applied in 4x4 pixel formats and rendered by the GPU. However, NVIDIA has trained small neural networks that can produce the desired pixel format and texture appearance at a fraction of the size of traditional texture compression formats. Instead of using gigabytes of VRAM for textures, NTC drastically reduces VRAM usage by emulating textures, allowing for either much lower VRAM consumption or significantly enhanced material appearance, depending on the game developer's goals. This enables games to incorporate much more complex scenery without any performance penalty, relying on NVIDIA's AI technology to handle the workload. Below is the demonstration of Tuscan Villa, which shows just how the scene looks.

If we take a look at the interior showcasing the tableware, you can see how NVIDIA's NTC actually pulls ahead of downscaled BCn textures to give a much better result, operating in the same 970 MB VRAM capacity.

Neural Texture Compression works by training a neural network to understand what a Texel, the base pixel of a texture map, looks like on a specific material. NVIDIA has trained these neural networks for nearly every game material, making them ready for real-world game deployment. This process is so refined that the output can either provide a more realistic version on top of the base texture layer that the game uses or maintain the same texture appearance for significant VRAM savings. For a full technical explanation and a look at what Neural Texture Compression does behind the scenes, you can watch the in-depth YouTube video about the technology.

63 Comments on NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB

#1 Very cool tech. Remember seeing it first proposed 7 years ago. Still waiting to see it in games. #2 Cool if true and it's actually functional like they say it is. With that said....

I can't wait for RTX 6090 4GB to be priced at $3k. #3

Event HorizonVery cool tech. Remember seeing it first proposed 7 years ago. Still waiting to see it in games.

Remember that NVIDIA has a literal supercomputer running for years to optimize DLSS and tech alike. So that is even cooler #4 Cool but hope it avoids texture blurring and ghosting like FG. If this means they are keeping 6070 at 12GB they are done. #5 They always fail to mention that the price of the memory savings is tanking performance in every single scene being rendered.

Putting neural networks invocation into the path of texture sampling is a mistake, because it's a task that needs to be done a lot and when using regular compression, its is very efficiently accelerated by specialised texture samplers. Textures are the right place to trade more memory usage for better performance.

This neural texture tech allows GPU vendor to skimp on memory and the advantage is that... FPS drops a lot. And battery life when gaming away from grid. #6

AleksandarKRemember that NVIDIA has a literal supercomputer running for years to optimize DLSS and tech alike. So that is even cooler

That is true. However, Intel and AMD are now doing the same for XeSS and FSR respectively. #7 Subtext for you the consumer: Expect to get 25-40% of the current RAM amount across the next GF line (while maintaining the the regular uplift of cost of course).
Less is more.

2500$-3000$ 12GB 6090 here we go :nutkick: #8

Ols-HolThey always fail to mention that the price of the memory savings is tanking performance in every single scene being rendered.

Putting neural networks invocation into the path of texture sampling is a mistake, because it's a task that needs to be done a lot and when using regular compression, its is very efficiently accelerated by specialised texture samplers. Textures are the right place to trade more memory usage for better performance.

This neural texture tech allows GPU vendor to skimp on memory and the advantage is that... FPS drops a lot. And battery life when gaming away from grid.

Neural texturing is actually faster because it requires fewer render passes. All the numbers are in the video that you didn’t watch before feeling the overwhelming need to post wrong information. #9 Something needs to compute these textures into existence, so they cut down the memory requirements and jacked up GPU requirements. 1,5 kW GPU's with even more melting power connector, here we come!

I know brute forcing through things isn't a solution but until this whole "lets use all memory in existence to power bullshit Ai" the memory was actually the cheapest component to increase its capacity. And by far the easiest. Just stack more of this shit on the PCB. But I guess NVIDIA's idea the entire time was drag all of us on bullshit cloud or make us all dependent on VRAM availability. Which is what they've been doing for years now offering unreasonably low amount of VRAM on graphic cards. 8GB just shouldn't even be a thing on something like RTX 5060. They should only come with 16GTB of VRAM by default.

Also new graphic card with more VRAM solves new games and old games in terms of improving performance. This neural bullshit needs to be specifically coded for the game. Meaning the shiny new RTX 7090 with just 8GB of VRAM because it'll all be this neural bs will be absolute ass with everything but games coded specifically for it. And seeing trends, they absolutely want to force this crap on us. #10 I don't have nerves to watch the video. Please educate me. Only boring static backgrounds are subject to reduction? #11

Dirt ChipSubtext for you the consumer: Expect to get 25-40% of the current RAM amount across the next GF line (while maintaining the the regular uplift of cost of course).
Less is more.

2500$-3000$ 12GB 6090 here we go :nutkick:

Well, nobody here values the RAM as being worth money by itself but rather by what you can do with it.

So I don't think that will happen unless Nvidia can get this technology to work on literally everything that has been released in the last five to ten years at least (perhaps a bit less, I am being rather cynical here) without need for the developers to do any implementation stuff.

And there's other stuff too like GPU compute/AI applications which need the VRAM, for which I'm not sure Nvidia would have the gall to say "you need to buy our 3x priced RTX pro-vis card for that, if you can't then get fucked" #12 Looks promising, but I don't think we can assume yet that it will work this wonderfully in real games across the board. Demos and controlled examples are one thing, broad adoption and consistent results are another. Still, since VRAM capacity is growing more slowly than we need, either much larger VRAM increases, or technologies like this are necessary to compensate. #13

Dirt ChipSubtext for you the consumer: Expect to get 25-40% of the current RAM amount across the next GF line (while maintaining the the regular uplift of cost of course).
Less is more.

2500$-3000$ 12GB 6090 here we go :nutkick:

Nah, 5090 is already 3000-4000. I think you want to say 6090 for 5000-6000? :) Lets be realistic. You probably guessed the VRAM tho :D #14 In theory, this is the kind of approach the industry needs. AI-related compute throughput keeps rising much faster than VRAM capacity, so better compression and neural asset handling could help balance that mismatch! #15 Not only 8GB VRAM is enough, the new standard will be 1GB VRAM for mid range! #16 seems games are not optimised. Did nvidia took on purpose something with a lot of noise and removed the noise? #17

phintsCool but hope it avoids texture blurring and ghosting like FG. If this means they are keeping 6070 at 12GB they are done.

Hopefully the next Gen cards will only need a fraction of that 12GB. That is the real goal. Bring down costs and improve images. #18 For everyone whining about VRAM, Nvidia isn't taking that away. This will require lots of memory bandwidth (like all AI tools), and wide memory buses mean lots of memory to go with them.

And why is nobody thinking about the savings in disk space? 90% of a modern game's size is game textures, most of it for objects you'll never even see up close. This can turn a 100GB game into a 25GB game. That's a huge win for consumers. #19

LastDudeALiveFor everyone whining about VRAM, Nvidia isn't taking that away. This will require lots of memory bandwidth (like all AI tools), and wide memory buses mean lots of memory to go with them.

And why is nobody thinking about the savings in disk space? 90% of a modern game's size is game textures, most of it for objects you'll never even see up close. This can turn a 100GB game into a 25GB game. That's a huge win for consumers.

Because it is easier to doompost about evil nGREEDia selling you 8GB GPUs (which they will defend otherwise) then it is to admit nVidia has made yet another technology that looks to be actually useful.

RejZoRSomething needs to compute these textures into existence, so they cut down the memory requirements and jacked up GPU requirements. 1,5 kW GPU's with even more melting power connector, here we come!

I know brute forcing through things isn't a solution but until this whole "lets use all memory in existence to power bullshit Ai" the memory was actually the cheapest component to increase its capacity. And by far the easiest. Just stack more of this shit on the PCB. But I guess NVIDIA's idea the entire time was drag all of us on bullshit cloud or make us all dependent on VRAM availability. Which is what they've been doing for years now offering unreasonably low amount of VRAM on graphic cards. 8GB just shouldn't even be a thing on something like RTX 5060. They should only come with 16GTB of VRAM by default.

Also new graphic card with more VRAM solves new games and old games in terms of improving performance. This neural bullshit needs to be specifically coded for the game. Meaning the shiny new RTX 7090 with just 8GB of VRAM because it'll all be this neural bs will be absolute ass with everything but games coded specifically for it. And seeing trends, they absolutely want to force this crap on us.

if nvidia wanted to force you onto cloud services, they never would have made 16GB GPUs int he first place, and would have never sold you anything over a 5060. #20

RejZoRSomething needs to compute these textures into existence, so they cut down the memory requirements and jacked up GPU requirements. 1,5 kW GPU's with even more melting power connector, here we come!

I know brute forcing through things isn't a solution but until this whole "lets use all memory in existence to power bullshit Ai" the memory was actually the cheapest component to increase its capacity. And by far the easiest. Just stack more of this shit on the PCB. But I guess NVIDIA's idea the entire time was drag all of us on bullshit cloud or make us all dependent on VRAM availability. Which is what they've been doing for years now offering unreasonably low amount of VRAM on graphic cards. 8GB just shouldn't even be a thing on something like RTX 5060. They should only come with 16GTB of VRAM by default.

Also new graphic card with more VRAM solves new games and old games in terms of improving performance. This neural bullshit needs to be specifically coded for the game. Meaning the shiny new RTX 7090 with just 8GB of VRAM because it'll all be this neural bs will be absolute ass with everything but games coded specifically for it. And seeing trends, they absolutely want to force this crap on us.

They won't reduce the memory capacity on the newer gens but there might be some stagnation which does suck. 32GB+ VRAM on mid-range cards would've been nice.

Their datacenter cards are pushing couple hundred GBs of VRAM in TB/s.

Interesting technology regardless of the business side of things.

TheinsanegamerNBecause it is easier to doompost about evil nGREEDia selling you 8GB GPUs (which they will defend otherwise) then it is to admit nVidia has made yet another technology that looks to be actually useful.

if nvidia wanted to force you onto cloud services, they never would have made 16GB GPUs int he first place, and would have never sold you anything over a 5060.

/jk (commented above, merged with this one)

#21 AI-Powered Cloud-VRAM incoming!
Get ready for your 2GB 6090! #22 Get ready for 1TB games, because 16K textures. #23 Arh all you clever people. But no 1 gb vram is not the new low standard. 128 mb of vram and 1 gb will be for the server gpus #24

Ols-HolThey always fail to mention that the price of the memory savings is tanking performance in every single scene being rendered.

Putting neural networks invocation into the path of texture sampling is a mistake, because it's a task that needs to be done a lot and when using regular compression, its is very efficiently accelerated by specialised texture samplers. Textures are the right place to trade more memory usage for better performance.

This neural texture tech allows GPU vendor to skimp on memory and the advantage is that... FPS drops a lot. And battery life when gaming away from grid.

I agree, although it might be possible to create an AI ASIC that can do a better job.

Running it on general AI cores is just trading cheap memory for expensive GPU die space as you pointed out. There's also the question of stability and AI artifacts. It has to be 100% stable and not cause a loss of fine detail as AI tends to do.

I'm not really sure why AI must be required for this on the consumer end. If the AI can better identify similar data patterns over traditional decompression, why can't it store it in a format that regular decompression units can use? Heck, even a new decompression ASIC would have been a much better alternative to using the general AI cores which are not going to be anywhere near as efficient. I suppose using AI on both ends in the easier solution to implement for Nvidia and it sells them more video cards but it almost certainly isn't the best option they could have come up with.

TheDeeGeeGet ready for 1TB games, because 16K textures.

That might be part of the plan. Make the memory usage untenable for anyone but those with the latest Nvidia cards. Everyone else with last gen and older cards will have to accept unoptimized massive textures when game devs only care about the NTC textures. It's a similar fear to one people have with DLSS 5.

RejZoRThis neural bullshit needs to be specifically coded for the game.

Not just for the game but per PBR material. They need to make a new AI model for each material according to Nvidia.

TheinsanegamerNBecause it is easier to doompost about evil nGREEDia selling you 8GB GPUs (which they will defend otherwise) then it is to admit nVidia has made yet another technology that looks to be actually useful.

Really depends on the performance overhead, the dev overhead, and the power consumption. Traditional decompression units are very efficient. You are basically hoping the bandwidth savings overcome the fact that Nvidia's general AI cores are vastly less efficient than an ASIC specifically designed for texture decompression. Nevermind the fact that the AI could have been used to compress the data into a format the ASICs can handle (or Nvidia could have updated it's texture decompression ASISs to handle a new data format). This may or may not be the future but given the facts, this appears more to be an attempt at something very much designed to benefit Nvidia and not the best way to go about it. #25 If trying to save VRAM, users would likely reject the feature if it comes at a performance cost especially if used on a large number of objects and textures where you are likely to run into tensor core bottlenecks. You can't really add an AI overhead and not expect a performance hit. A better solution is for Nvidia to stop skimping on VRAM.
If an AI penalty is to happen, then it should be used for something that cannot be done with traditional textures in a practical sense. For example, In a game, you likely wouldn't use a 5+GB megascan texture for a wall so that it can be wall hugging friendly such that you can get close enough to see near microscopic detail. But imagine if there were a neural texture that could generate new detail (rather than just compression) that can take over when you get close enough, and then the AI model gives near infinite detail to the surface. Add your own comment

출처: Techpowerup.com

원문 게시일: 2026-04-04T17:11:57Z

이 글은 NewsAPI를 통해 자동으로 수집된 콘텐츠입니다. 원문은 위 링크를 통해 확인하세요.

NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB - TechPowerUp

요약

본문

Related News

63 Comments on NVIDIA's Neural Texture Compression Cuts VRAM Use From 6.5 GB to 970 MB