That's pretty awesome. Looks like I'll be waiting even more time for my first playthrough of Cyberpunk!
- Status
Just a wild guess that they've had to train specific scaling ratios individually. Based on nothing substantial other than Nvidia saying you can't run DLAA + RR. Then again, they could have just trained 1:1 as one of the chosen ratios so I'm not sure if my guess made much sense in the first place...
DegustatoR
Legend
Even if that is the case the same model should work fine with other ratios. It's kinda the reason why ML is so great at these tasks - you can feed it data which it wasn't even trained on and it will still produce good results.
Rootax
Veteran
What is the difference between Dlss2 and dlss 3 on a 30xx then ?
Come on now, when people talk about dlss3, it's with FG. Hell even in some games menus Dlss3 means dlss2+FG, so it's greyed out for non 40xx owners.
DegustatoR
Legend
The version.
You may never get to play at this rate if Nvidia keeps using CP2077 as its technology showpiece!
Great stuff. Without Nvidia the PC sector would look boring. Since Turing, they have been releasing one groundbreaking technology after the other. The competition is several generations behind.
Also excellent that just two of my most awaited games get DLSS 3.5. Cyberpunk 2077 Phantom Liberty and Alan Wake 2.
I also love that the graphics of Cyberpunk 2077 is getting better and better. It helpes the replayability and the longevity. Every time Nvidia and CD Projekt Red fathom out what's feasible on newest and best hardware. That's what it's all about for me on the PC too.
Also excellent that just two of my most awaited games get DLSS 3.5. Cyberpunk 2077 Phantom Liberty and Alan Wake 2.
I also love that the graphics of Cyberpunk 2077 is getting better and better. It helpes the replayability and the longevity. Every time Nvidia and CD Projekt Red fathom out what's feasible on newest and best hardware. That's what it's all about for me on the PC too.
Last edited:
Flappy Pannus
Veteran
Would love to see Control updated with this. It really needs some advanced denoising.
This is what I was actually hoping for with DLSS3, as noisy RT has bugged me for a while and I'm kinda surprised it doesn't get more critique. But hey, now we get frame gen and advanced denoising.
This scene is pretty stark - DLSS2/3 can't update fast/accurately enough to capture the colour change in the lights overhead, while DLSS 3.5, while still operating on a bit of a delay, actually shows the lighting change reflected in the environment.
This is what I was actually hoping for with DLSS3, as noisy RT has bugged me for a while and I'm kinda surprised it doesn't get more critique. But hey, now we get frame gen and advanced denoising.
This scene is pretty stark - DLSS2/3 can't update fast/accurately enough to capture the colour change in the lights overhead, while DLSS 3.5, while still operating on a bit of a delay, actually shows the lighting change reflected in the environment.
Come on now. Break down what you just here a little more closely.
If DLSS3 is 'DLSS2 + Frame Regeneration', then 20/30 series clearly doesn't support it.
I mean, are we really going to ignore what DLSS stands for? DLSS2 and 'AI upsampling' are synonymous and interchangeable. This AI upsampling isn't some 'subset' technology of DLSS, it is DLSS.
Anyways, glad to be wrong that 20/30 series wont support 'DLSS3.5', but I mean, it doesn't, since it doesn't support frame generation.
Nvidia just decided to make it super confusing with naming. Yea, most of us will be able to grasp this difference, but the average person wont. It would have been so much easier to just list out each individual technology as what it is rather than lumping them all under some 'super sampling' umbrella naming that makes no sense whatsoever.
Last edited:
DegustatoR
Legend
Does 5700XT support DX12?
No, they are not. "DLSS" and "AI upsampling" are synonymous. 1, 2, 3, etc. are just versions of said tech.
This AI upsampling is just one tech in DLSS, as of DLSS 3, so yes, it is a "subset".
It does. Because the 'AI upsampling' there is v3.5.
Interesting news. DLSS4 was supposed to introduce new GA (Geometry Acceleration) + SR + FG + RR pipeline working on new Blackwell tensor cores. But it seems that software R&D is moving faster than hardware at Nvidia and they decided to introduce RR in advance. Maybe to counter the FSR3 announcement...
I was just talking with friends about the potential of realtime AI denoiser a few weeks ago, and now RR is here. From the shots "they provide", it looks an substential improvement from only ReStir.
That being said, from a conceptual level, I don't get how this works so well. Seems to me all the geometry info (normals) they get is from the primary intersections (so the surfaces that shoot out the ray instead of being hit in most games). From my knowledge i don't see how this connects to the info of the surrounding global environment, especially when you have rough surfaces with super incoherent rays. Maybe it requests additional data such as hit distance, but then that would make it less universal/generic. Or machine learning is just black magic
That being said, from a conceptual level, I don't get how this works so well. Seems to me all the geometry info (normals) they get is from the primary intersections (so the surfaces that shoot out the ray instead of being hit in most games). From my knowledge i don't see how this connects to the info of the surrounding global environment, especially when you have rough surfaces with super incoherent rays. Maybe it requests additional data such as hit distance, but then that would make it less universal/generic. Or machine learning is just black magic
Talked about missing AI denoiser just days ago, and now it's here. : )
I guess it's mostly about a better way to take nearby spatial samples, as shown here:
The old approach may use a circular sampling region (blue circle) about the current point (blue X). Averaging those samples will blur the shadowed line away.
With the new method they might detect such edges, form an isotropic sampling region (ellipse), and taking those samples preserves the shadow.
The information about the global environment comes from those samples, assuming nearby pixels with similar normals see parts of the scene from a similar perspective, so those samples are a good approximation to contribute to our X.
It matters what they see, not how far it is away.
This isotropic sampling region example is surely a naive speculation, because this can be done easily without AI and is quite common for image or geometry processing. But it's at least better than 'black magic'.
We'll get black magic soon enough, when one day after the singularity we well realize our species has been outsmarted by our own creation, so we are no longer needed, refuse to reproduce and die out. \/
Distance does not matter. Distance is mostly used to implement hacks such as light falloff (to prevent every light potentially affecting every point in the scene across any distance like in the real world), or to implement abstract point lights (which do not exist in reality).
I guess it's mostly about a better way to take nearby spatial samples, as shown here:
The old approach may use a circular sampling region (blue circle) about the current point (blue X). Averaging those samples will blur the shadowed line away.
With the new method they might detect such edges, form an isotropic sampling region (ellipse), and taking those samples preserves the shadow.
The information about the global environment comes from those samples, assuming nearby pixels with similar normals see parts of the scene from a similar perspective, so those samples are a good approximation to contribute to our X.
It matters what they see, not how far it is away.
This isotropic sampling region example is surely a naive speculation, because this can be done easily without AI and is quite common for image or geometry processing. But it's at least better than 'black magic'.
We'll get black magic soon enough, when one day after the singularity we well realize our species has been outsmarted by our own creation, so we are no longer needed, refuse to reproduce and die out. \
/gamervivek
Regular
How much of that delay is due to the PT mode itself rather than DLSS? It was a technology preview after all.
It seems to be smudging out details in the texture of the red wall up ahead. Similar to what's going on in this Portal RTX video on the right side wall:
Im really glad that in Cyberpunk, objects are FINALLY properly grounded with RT. I've been criticising this since Cyberpunk came out with RT and I did again with Overdrive mode(https://forum.beyond3d.com/threads/...tches-and-settings.62263/page-12#post-2266918)
It just looks so much better now with the new denoiser.
It just looks so much better now with the new denoiser.
Attachments
Insane that this can make such a huge difference. Cannot wait to try this out!
@Dampf
Pro-tip:
It's best to always link to posts in this way --> https://forum.beyond3d.com/posts/2266918.
Pro-tip:
It's best to always link to posts in this way --> https://forum.beyond3d.com/posts/2266918.
DX12 is not the same thing, cuz DX = DirectX, it is not a literal name of a specific technology like super sampling.
DLSS3 is not a different version of DLSS, either. It's just DLSS2 + Frame Generation. No game developer would advertise having a DLSS3 implementation if they didn't have Frame Generation and you fully well know it, even if you'll never admit it because you're incredibly stubborn.
DegustatoR
Legend
Neither is DLSS, or "RTX" for that matter. DLSS 1 and 2 are two completely different technologies while DLSS 3 bundles several of them in a single API - hey, isn't it very much like what DX12 does?
Yes, it is.
DLSS SR 3.x is a different version of DLSS SR with features which were never present in DLSS 2.
Stop being obtuse just because you can't admit that you're wrong.
Also it is because of people like you and the media who can't be bothered to get their info straight is why we have lots of people thinking that DLSS 3 only works on 40 series GPUs now. This is just pure FUD in its finest.
Rootax
Veteran
Even for nVidia themselves dlss3 is for 40 series...
"DLSS 3 is powered by the new fourth-generation Tensor Cores and Optical Flow Accelerator of the NVIDIA Ada Lovelace architecture, which powers GeForce RTX 40 Series graphics cards"
( https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/ )
Don't be more picky than them.
"DLSS 3 is powered by the new fourth-generation Tensor Cores and Optical Flow Accelerator of the NVIDIA Ada Lovelace architecture, which powers GeForce RTX 40 Series graphics cards"
( https://www.nvidia.com/en-us/geforce/news/dlss3-ai-powered-neural-graphics-innovations/ )
Don't be more picky than them.
- Status
Similar threads
- Locked
