• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

AMD finally supports Neural Texture Compression!

are closed said:
Don't get me wrong, I'm all for positive vibes, but I don't see many developers going back and patching existing games to lower vram footprint.
Click to expand...
I mean games that are midway through active development and haven't yet launched, not already launched titles.
Having the tech backported to existing launched titles would be some next-level shit I never even considered.
 
Lakados said:
The 6000 series lacks matrix multiplication capabilities, so those cards just can't handle the math.
Click to expand...
So your saying no console till maybe the end of next year, but realistically likely the year after that can take advantage of this?
Ya unfortunately this is meaningless for now. No one is going hard (re)developing games for a 2030 hardware landscape.
 
ChadD said:
So your saying no console till maybe the end of next year, but realistically likely the year after that can take advantage of this?
Ya unfortunately this is meaningless for now. No one is going hard (re)developing games for a 2030 hardware landscape.
Click to expand...
PS6 has the Neural Array Cores, which provide the needed mathematical functionality, and likewise the next Xbox CPU 'Magnus' has the needed capabilities as well, and both of those were scheduled for 2027, but the market may push them back to 2028...
Sony developed their own form of Neural Texture Compression as part of Project Amethyst for the console release and was showing it off on demo hardware in October of 2025.

I am just trying to be hopeful about something for once.
 
Lakados said:
PS6 has the Neural Array Cores, which provide the needed mathematical functionality, and likewise the next Xbox CPU 'Magnus' has the needed capabilities as well, and both of those were scheduled for 2027, but the market may push them back to 2028...
Sony developed their own form of Neural Texture Compression as part of Project Amethyst for the console release and was showing it off on demo hardware in October of 2025.

I am just trying to be hopeful about something for once.
Click to expand...

Fair. I just don't see it. :) The consoles are for sure getting pushed back to 2028, and even then unless the AI bubble completely bursts... they aren't going to move. This is nice, it just doesn't change much. Had AMD said we found a way to make this work on current gen console hardware. That would be huge. Being tied to future hardware likely to be selling at prices that hamper sales even in 2028. I think this is still tech that the majority of gamers aren't going to have access to until the 2030s.
Even if AMD has some Uber mid range RDNA5 card they can sell at 9070 MSRP type pricing. Its going to be ruff if Sony and MS are just not shipping a new gen, or launching base models at pro prices.

I appreciate your trying to be upbeat. :)
 
Lakados said:
I mean games that are midway through active development and haven't yet launched, not already launched titles.
Having the tech backported to existing launched titles would be some next-level shit I never even considered.
Click to expand...

For PS5 support (and older 6000-7000 series AMD gpu, pascal and possibly low end turing/ampere for performance reason) you would need to do all the traditional BC7 type work already and make it fit on 8GB vram popular gpu on the computer side with it, would be good money to make an different version of the asset/render path and would not save place on the download/install size...

I imagine this will be small title...PS6 generation exclusive in the earliest for big stuff with a big risk.

edit: I guess they could make first launch slower (like pre shader compilation and textures from Neural network decompression) for older generation GPUs and use this already.
 
Last edited:
DukenukemX said:
I'm betting on MESA will add it for 6000 series cards. Is AMD the first to implement Neural Texture Compression? Been hearing Nvidia doing it for some time now.
Click to expand...
been going on for a while, it is recent that Nvidia suppor the DX12/vulkan general way to do it, the cuda NR has been there for a while, https://github.com/NVIDIA-RTX/RTXNTC/commits/main/ (february 6, 2025, first commit), RTX Neural Texture Compression

Issue is, inference cost can be big if you do not have the hardware for this, (always possible to work, but framerate cost is too big), that a saving vram and bandwith in exchange of compute affair.

But there is multiple way to use this, you can at load time transform the very neural compressed into regular BC7 type texture (and you can choose resolution-quality that will match the gpu vram dynamically) and use them in a regular way after that or in real time per pixel, so there could be something to do for old hardware.

Inference on load could work on those older gpus, that would make a single path dev wise, save on download game size and can make quality of texture adjust well for a specific scene/vram capacity, but will not save vram during gameplay.
 
DukenukemX said:
I'm betting on MESA will add it for 6000 series cards. Is AMD the first to implement Neural Texture Compression? Been hearing Nvidia doing it for some time now.
Click to expand...
They are the last, Intel got it working for ARC in November of 2025.

The biggest issue is we currently have at least 4 competing standards for how to do this, and I am unsure about compatibility.

Nvidia has their "open" standard.... https://github.com/NVIDIA-RTX/RTXNTC
AMD now has theirs. https://github.com/GPUOpen-LibrariesAndSDKs/MiniDxNn
Intel also has an open one too. https://github.com/GameTechDev/TextureSetNeuralCompressionSample

And Sony was reportedly working with AMD on theirs for the PS6, but Sony does their own thing when it comes to texture compression so who knows what that looks like.

The only company that supposedly has a brand-agnostic, AI-driven texture compression technology working is Epic.
Epic has it in the Experimental section of the UE5 plugins; it's in the Neural Rendering package.
https://dev.epicgames.com/documentation/unreal-engine/neural-post-processing-in-unreal-engine
Some of the demos I have seen reduce VRAM from ~7GB down to less than 1... So.... Yeah, that's crazy, but also pretty sandboxed, so who knows what that looks like in the wild.

Regarding MESA and getting support for it on the RX 6000 series I don't know if it is worth the effort.

The RX 6900XT, going all out, is capable of 1/5'th the number of Matrix Multiplications per second that the RX 9060 XT can manage. But the 9060XT does it with a sub-processor, and it doesn't have to engage the rendering cores to manage that workload.

So I am sure that it is technically feasible, but I fear it would be one of those situations that succeeds by reducing the game to single-digit FPS numbers.
 
Last edited:
LukeTbk said:
been going on for a while, it is recent that Nvidia suppor the DX12/vulkan general way to do it, the cuda NR has been there for a while, https://github.com/NVIDIA-RTX/RTXNTC/commits/main/ (february 6, 2025, first commit), RTX Neural Texture Compression

Issue is, inference cost can be big if you do not have the hardware for this, (always possible to work, but framerate cost is too big), that a saving vram and bandwith in exchange of compute affair.

But there is multiple way to use this, you can at load time transform the very neural compressed into regular BC7 type texture (and you can choose resolution-quality that will match the gpu vram dynamically) and use them in a regular way after that or in real time per pixel, so there could be something to do for old hardware.

Inference on load could work on those older gpus, that would make a single path dev wise, save on download game size and can make quality of texture adjust well for a specific scene/vram capacity, but will not save vram during gameplay.
Click to expand...

Yep, it's the classic storage vs compute trade-off in Computer Science 101.

I invision that looking like shader caching when a game starts up. However, I could also see a downloaded game at 50GB exploding to 350GB after said texture expansion, to support older video cards.
 
You must log in or register to reply here.
Back
Top