NVIDIA Introduces Revolutionary Neural Texture Compression for Material Textures

Published by

Click here to post a comment for NVIDIA Introduces Revolutionary Neural Texture Compression for Material Textures on our message forum
https://forums.guru3d.com/data/avatars/m/263/263841.jpg
Okay so how is this going to be implemented? If it's closed source then game developers are going to have to pack all the regular textures in with the game anyway to account for non supported cards on 3 teams plus the consoles. So game installs are going to be bigger again, some developers won't bother implementing it so nVidia cards need to have large enough VRAM to cover those cases at which point why are you sacrificing cycles for a feature that doesn't really have a benifit
https://forums.guru3d.com/data/avatars/m/243/243189.jpg
Neo Cyrus:

It says it's done with matrix multiplication (think of a matrix as a grid of sets of data containers; a tensor is higher dimensional matrix), so it sounds like it'll need tensor cores/whatever specific cores for acceleration. nVidia won't allow it to work with Intel's hardware, and I don't remember if AMD's GPUs have any such hardware, I don't think they do.
Funnily enough I stumbled across this information whilst playing with this: https://github.com/nod-ai/SHARK. The instinct industry cards have more matrix support function with dedicated matrix processing units (https://gpuopen.com/learn/amd-lab-notes/amd-lab-notes-matrix-cores-README/) The 7000 series has something similar to tensor cores rather than tensor cores themselves, more generic AI accelerators which amongst other things can perform matrix multiplication functions (https://gpuopen.com/learn/wmma_on_rdna3/), but no idea about their actual performance in this area compared to what Nvidia claims. If stable diffusion output is anything to go by, it's pretty good.
https://forums.guru3d.com/data/avatars/m/132/132389.jpg
moo100times:

Funnily enough I stumbled across this information whilst playing with this: https://github.com/nod-ai/SHARK. The instinct industry cards have more matrix support function with dedicated matrix processing units (https://gpuopen.com/learn/amd-lab-notes/amd-lab-notes-matrix-cores-README/) The 7000 series has something similar to tensor cores rather than tensor cores themselves, more generic AI accelerators which amongst other things can perform matrix multiplication functions (https://gpuopen.com/learn/wmma_on_rdna3/), but no idea about their actual performance in this area compared to what Nvidia claims. If stable diffusion output is anything to go by, it's pretty good.
From what I've heard (I didn't verify it, so take it with a grain of salt) AMD's 7K series cards don't have dedicated hardware for matrix multiplication, instead they're run on shared units... Edit: I just asked a completely unreliable source to verify (the best kind of source), and apparently I remembered wrong, the RT units are built into the texture units for AMD so that's what I was thinking of, the matrix multiplication is apparently done on the shaders. So yeah, no dedicated tensor/mm hardware on AMD cards.
https://forums.guru3d.com/data/avatars/m/273/273822.jpg
So my 12gb fubar GPU is now as good as an AMD 24gb GPU? Cheers.
https://forums.guru3d.com/data/avatars/m/273/273822.jpg
Undying:

And convince nvidia to ship graphics cards with memory to have room for the textures.
For someone who's been beating the "VRAM GIMEEEEE MOAR VRAM" for quite a few years, it's quite funny to see you using a GPU that "doesn't have enough VRAM".
https://forums.guru3d.com/data/avatars/m/243/243189.jpg
Neo Cyrus:

From what I've heard (I didn't verify it, so take it with a grain of salt) AMD's 7K series cards don't have dedicated hardware for matrix multiplication, instead they're run on shared units... Edit: I just asked a completely unreliable source to verify (the best kind of source), and apparently I remembered wrong, the RT units are built into the texture units for AMD so that's what I was thinking of, the matrix multiplication is apparently done on the shaders. So yeah, no dedicated tensor/mm hardware on AMD cards.
Agreed. It's only the instinct cards that have dedicated Matrix Core Processing Units.
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
metagamer:

For someone who's been beating the "VRAM GIMEEEEE MOAR VRAM" for quite a few years, it's quite funny to see you using a GPU that "doesn't have enough VRAM".
You might be right. 12gb vram is the barely enough for 1440p these days.
data/avatar/default/avatar08.webp
metagamer:

For someone who's been beating the "VRAM GIMEEEEE MOAR VRAM" for quite a few years, it's quite funny to see you using a GPU that "doesn't have enough VRAM".
That's nothing. In his younger Ryzen 1 days he was screaming GIMME 8+ CORES OR DEATH. .. That was before he... "upgraded" to 6 cores :D
https://forums.guru3d.com/data/avatars/m/56/56686.jpg
if gona be locked behind Nvidia tech, that it will probably never really take off, much like Gsync seeing most monitor are gsync compatible and not actual have module in it. would be great if becomes industy standard, gona have to wait till amd does it verison of and dont lock it behind just there card
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
Noisiv:

That's nothing. In his younger Ryzen 1 days he was screaming GIMME 8+ CORES OR DEATH. .. That was before he... "upgraded" to 6 cores :D
Games didnt took the route we hopped for and still heavily singlethread limited. It is what it is 😀
data/avatar/default/avatar06.webp
The issue is not that AMD/Intel do not have the required hw for fast matrix multi. It's that Nvidia's readily available, highly optimized matrix routines run on top of CUDA software. This particular implementation from the whitepaper can be run on PyTorch but would require 18 GB and still be 10x slower. When Jensen says they're first and foremost a software company, he's proly right.
Undying:

Games didnt took the route we hopped for and still heavily singlethread limited. It is what it is 😀
They're not heavily ST. That's STALKER. But games are not, nor will they ever be, infinitely and perfectly threaded. And there will always be a CPU pig or two, or a bad launch, when MOAR CORES will be left hanging, while the single-thread fast CPU will be able to drive through the code.
https://forums.guru3d.com/data/avatars/m/273/273822.jpg
Undying:

You might be right. 12gb vram is the barely enough for 1440p these days.
It's absolutely fine for any game I've played.
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
metagamer:

It's absolutely fine for any game I've played.
I know you like Diablo so we'll talk about it when it releases. Beta was not fine.
https://forums.guru3d.com/data/avatars/m/273/273822.jpg
Undying:

I know you like Diablo so we'll talk about it when it releases. Beta was not fine.
Dude, I told you about the beta already. It was clearly bugged on high res textures, regardless of what VRAM your GPU had. The textures looked mediocre at best anyway, plus there's no draw distance in an isometric top down ARPG, it's not like the game will need 16GB VRAM. This will be fixed on release or shortly after and anyone with 8GB VRAM GPUs will run it just fine. Btw, high res vs no high res textures in the beta looked nigh on identical.
https://forums.guru3d.com/data/avatars/m/198/198862.jpg
metagamer:

Dude, I told you about the beta already. It was clearly bugged on high res textures, regardless of what VRAM your GPU had. The textures looked mediocre at best anyway, plus there's no draw distance in an isometric top down ARPG, it's not like the game will need 16GB VRAM. This will be fixed on release or shortly after and anyone with 8GB VRAM GPUs will run it just fine. Btw, high res vs no high res textures in the beta looked nigh on identical.
I hope you are right and we'll also see about that rt support and how it handles it.
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
4gb 64bit 5060 ti confirmed ? :P
https://forums.guru3d.com/data/avatars/m/132/132389.jpg
Noisiv:

The issue is not that AMD/Intel do not have the required hw for fast matrix multi.
Intel Arc GPUs have dedicated matrix units.
https://forums.guru3d.com/data/avatars/m/121/121558.jpg
Genuinely impressive tech for non-affordable graphics cards. The minority will benefit unless NVIDIA one day can realize that more people buying their cards = making more money.
https://forums.guru3d.com/data/avatars/m/40/40458.jpg
Zenoth:

Genuinely impressive tech for non-affordable graphics cards. The minority will benefit unless NVIDIA one day can realize that more people buying their cards = making more money.
85% market share is the minority?
https://forums.guru3d.com/data/avatars/m/238/238382.jpg
Nvidia > Creates problem Nvidia > Creates solution Makes me think of what "live service" video games do to make people spend money.