NVIDIA’s Neural Texture Compression for material texture compression

yinzerniner said:
Maybe Nvidia and Amd WERE RIGHT! We don’t need more than 8gb of vram


View: https://youtu.be/kQCjetSrvf4?si=URLMLvQscndct2nq
Click to expand...

The problem for AMD is that this tech makes extensive use of Matrix Multiplication for the decompression process.
Only the RDNA4 series would be able to handle this at a decent speed, which by the time we see it in the wild may not be an issue, but it will work on all the Nvidia RTX series of cards.
 
yinzerniner said:
Maybe Nvidia and Amd WERE RIGHT! We don’t need more than 8gb of vram


View: https://youtu.be/kQCjetSrvf4?si=URLMLvQscndct2nq
Click to expand...

I think this would likely have to be implemented on a per game basis. Textures need to be compressed in memory before they're loaded in a game into this particular format (it's a lot different than loading a jpeg or png). ASTC generally already does a pretty good job and is better than standard block, so this is kind of tricking people a bit into thinking this is better than it is tbh. There's also a ton of techniques to deal with vram usage and texture allocation in both unity and unreal. Beyond that I'm not sure what custom engines do, but you can be pretty clever with compression, steaming, and texture atlasing.
 
I wonder if it is youtube compression or if the tech changed in the last 3 months, because without DLSS (even with TAA) it was a bit noisy before
 
LukeTbk said:
I wonder if it is youtube compression or if the tech changed in the last 3 months, because without DLSS (even with TAA) it was a bit noisy before
Click to expand...
They did, they completely changed how they encode their streams.
Basically, if you aren't on premium, then regardless of what you select for your streaming quality, you are watching something that is some form of upscaled 720p.
 
Lakados said:
They did, they completely changed how they encode their streams.
Basically, if you aren't on premium, then regardless of what you select for your streaming quality, you are watching something that is some form of upscaled 720p.
Click to expand...
No i meant the texture compression tech, just compiled the updated version and it is still a bit noisy on TAA around high frequency change (yellow logo for example)
 
LukeTbk said:
No i meant the texture compression tech, just compiled the updated version and it is still a bit noisy on TAA around high frequency change (yellow logo for example)
Click to expand...
Oh, no clue, but yeah, YouTube changed from AV1 to VP9... people notice

I can't comment on the demo you're working with.
 
Lakados said:
The problem for AMD is that this tech makes extensive use of Matrix Multiplication for the decompression process.
Only the RDNA4 series would be able to handle this at a decent speed, which by the time we see it in the wild may not be an issue, but it will work on all the Nvidia RTX series of cards.
Click to expand...
This feature is not going to save 2000 series RTX cards, in terms of playing games released in 2027 or 2028.
This is a lot like things like Tesselation. AMD had in 3 generations of cards before it mattered. By the time it did the first couple generations could "use" it, but not really. And Nvidias first generation was as capable as AMDs current thing.
The Matrix units in 2000 cards are of no concern as they won't be capable of performing at a level required anyway. Pretty much the same deal with 3000 and 4000.
The cards that actually are rocking 8gb... even cards like 5060 with their 144 tensor cores, those cards only have so many cores to start handling all these features Texture compression AI, Upscaling AI, Ray Tracing AI. Maybe you can use the fancier textures but its not going to magically enable you to start flipping everything to crazy settings either. Same issue of course for AMD, the cards were this would actually help like the 9060 8gb are also light on silicon.
 
ChadD said:
ven cards like 5060 with their 144 tensor cores, those cards only have so many cores to start handling all these features
Click to expand...
if they can fit on 4 bits models that support accumulating something like a 5060 has still impressive raw power, a bit over 300/600 FP4 AI "tops", that higher than even a RTX 3090 peak int8 tops capacity and a bit more than twice its FP8 tops performance. In a world where some products are proud to announce they have 40 tops available. 5060 is about a 3090, but in 4 bits it is twice the card, way higher FP8 has well.

Looking at the code intel demo is running FP16 and NVIDIA with int8/fp8 right now.

By the time this become popular (if it ever do, what happened to directstorage on the gpu for example... it require many future things to fall a certain way including extra tensor core/inference capacity not ending up being used on something else instead ), 2000 series/3000 series would be quite old a bit like Pascal today, strong 8 bits generation like the 4000-5000 series would be at best the oldest but still recent enough to matter and probably what they would aim for if not 5000 and up, say if PS6 type of title start to use it that would be around what 2028-2029 ? (because if the ps5 is capable the 2000 series would and we can doubt we will not see big budget game made that are not made to run well on a PS5 before the PS6 is mature)

It could end up be more aiming at making 2010s Pixar level of asset quality being able to fit on a downloadable file in a reasonable time that fit on the small ssd people seem we will still have and able to run on the 24-32GB of ram console, 16GB+ gpu will have, more than try to make ancient 8GB gpu work well.
 
Last edited:
LukeTbk said:
if they can fit on 4 bits models that support accumulating something like a 5060 has still impressive raw power, a bit over 300/600 FP4 AI "tops", that higher than even a RTX 3090 peak int8 tops capacity and a bit more than twice its FP8 tops performance. In a world where some products are proud to announce they have 40 tops available. 5060 is about a 3090, but in 4 bits it is twice the card, way higher FP8 has well.

Looking at the code intel demo is running FP16 and NVIDIA with int8/fp8 right now.

By the time this become popular (if it ever do, what happened to directstorage on the gpu... ), 2000 series/3000 series would be quite old a bit like Pascal today, strong 8 bits generation like the 4000-5000 series would be at best the oldest but still recent enough to matter and probably what they would aim for if not 5000 and up, say if PS6 type of title start to use it that would be around what 2028-2029 ? (because if the ps5 is capable the 2000 series would and we can doubt we will not see big budget game made that are not made to run well on a PS5 before the PS6 is mature)

It could end up be more aiming at making 2010s Pixar level of asset quality being able to fit on a downloadable file in a reasonable time that fit on the small ssd people seem we will still have and able to run on the 24-32GB of ram console, 16GB+ gpu will have, more than try to make ancient 8GB gpu work well.
Click to expand...
The thing I wonder is whether offloading the generation of textures from rastered images into storage vs increased compute power is worse or better for overall system efficiency as well as chip optimization. It seems like the chips themselves will need to be of higher quality and compute power, who would obviously mean lower yields which is gonna be an issue regardless as they go down from 4nm. In the meantime gddr7 is costly now but will only get cheaper as manufacturing streamlines. So if this is implemented the chip supply available will go down even further than now, with AI demand only rising.

Granted AI and other machine learning uses also require all the VRAM, but would this feature also reduce the VRAM load for those uses as well?

Lot of questions, not enough answers but cool to see how it progresses regardless
 
yinzerniner said:
The thing I wonder is whether offloading the generation of textures from rastered images into storage vs increased compute power is worse or better for overall system efficiency as well as chip optimization. It seems like the chips themselves will need to be of higher quality and compute power, who would obviously mean lower yields which is gonna be an issue regardless as they go down from 4nm. In the meantime gddr7 is costly now but will only get cheaper as manufacturing streamlines. So if this is implemented the chip supply available will go down even further than now, with AI demand only rising.

Granted AI and other machine learning uses also require all the VRAM, but would this feature also reduce the VRAM load for those uses as well?

Lot of questions, not enough answers but cool to see how it progresses regardless
Click to expand...
The short answer is: “yes it is”.
The long answer is: <multiple long paragraphs overly simplifying the inefficiencies of the render pipeline as object density increases and resolutions increase go up, goes here>
 
You must log in or register to reply here.
Back
Top