Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[pjbliverpool]

The PS5 is categorically and objectively outperforming the 2060S by a substantial margin.

A lot of this thread has discussion from many folks on why the we should expect the PS5 to outperform it. Due to tflops, etc. And why it shouldn't be considered the as PS5 punching above its weight, because the 2060S is in a lower weight class. My impression of reading the thread is that most are in agreement, just a bit of arguing over the semantics.

Are you talking with DLSS or without now? Without DLSS then yes of course the PS5 should outperform a 2060/S in a non-RT game. I'm one of the people who's said that already in this thread several times. But I've not seen anyone making the same claim about a 2060S while using DLSS other than yourself. Given that DLSS can literally more than double performance in some cases I'd say it'd be unwise at best to make such a generalised statement.

But even in reference to this specific game, if that is your claim I'm just not seeing how you're reaching it. There's no common frame of reference and there's no reliable way to to equate extra resolution to performance loss in specific percentage terms.

I'll perform all of the calculations for you and break down all of your points later. I need to wait for a slow day at work to write everything out.

I'm certainly curious to see how you're reaching this conclusion. In particular how you will explain that the PS5 in either quality or standard mode needs to drop as low as 1440p at times in order to maintain 60fps while the DLSS output on the 2060 is running with 125% more resolution at 4K in those instances and goes no lower than 41fps (very briefly) in the DF video. So that's essentially 41fps at 4K vs 60fps at 1440p. How do you conclude from that that one is substantially better than the other? And that's just comparing to the 2060, not the 2060S.

In fact to look at this in a much simpler way, in the DF video (time stamped below) Alex points out that the performance uplift in his sample scene between native 4K and DLSS Quality and Performance modes are 31% and 50% respectively.

50% more performance than a 2060 according to TPU puts you comfortably above 2080S performance while 50% more performance than a 2060S put you between an RTX3070 and a RX 6800. Is it your assertion that the PS5 is "objectively outperforming the RTX 3070 by a substantial margin"?

Nvidia's own benchmarks corroborate that performance uplift here (note the 2060 slightely outperforming the 3060Ti with DLSS(P) activated.

 

[neckthrough]

I can't follow your logic.
How do you justify allowing one scaled guessworking technology to be counted as "approximation", but discard other scaled technologies?
Checkerboarding is just as much approximation as DLSS is even if it's IQ is in many ways worse.

Perhaps you misread my post. I said exactly what you are saying here -- they are all approximations, and so it's fair to compare them. I did add the caveat that in some cases if one approximation is shown to be consistently inferior to another, then that observation must be factored into the discussion. But the discussion is legitimate.

Counting native as approximation is just BS it was considered the ground truth (TAA or not) in game IQ discussions

There's nothing to be "considered" here, this is not a matter of opinion. Native is not ground truth, as anyone trained in computer graphics and/or signal processing will tell you. Even the 64x supersampled render used to train DLSS is an approximation of ground truth, but it's a much closer approximation than native.

until the day DLSS came and certain people started forcing it as equivalent despite it's clear issues (even with DLSS 2.0) because "it sometimes looks as good or even better" - none of which changes the fact it's upsampled guesswork with it's own issues and drawbacks.

100% agree that it's upsampled guesswork with its own issues and drawbacks. I also agree that it sometimes looks as good or even better than native or native+TAA. Once again, remember that DLSS is *not* trained to target a flawed native render -- if it did, it would, by construction, be consistently worse than native.

The moment you allow one scaling and/or guessworking method in, you need to allow them all and it becames a subjective playingfield.

Again, yes, but once again, caveat applies that if one approximation is shown to be consistently inferior to another then that needs to be called out.

Counting DLSS as native equivalent is just intellectual dishonesty, either you allow all and come up with your own personal IQ/FPS-grading system, or you allow none.

It's obviously not objectively equivalent -- the pixels are different. Subjectively, what most people seem to see is that in many implementations it trades blows with native. Therefore, they see them as equivalent-enough IQ-wise to use performance as the judging criterion. Obviously you disagree, and that's fair.

 

[Scott_Arm]

For you, it could look better than native, but that's a subjective opinion, not objective fact based on some established ruleset.

I mean what was considered and called native before DLSS and it's promotors came marching in claiming the native should suddenly be some hypothetical infinitely supersampled version of the image that could never really exist in a game.
So no, not just what I have called "native".

You're basically at the point of semantic arguments now. "Native" rendering is highly under-sampled, which is why we have aliasing and shimmering that needs to be removed in the first place. So all of these techniques like DLSS, TAA etc exist to try and improve a low sample image, and they all have different strengths and weaknesses. There's no purely objective way to evaluate them against each other, because it'll come down to personal preferences over blur, sharpness, ghosting and other artifacts.

Also, frame rate is an image quality problem. You can't evaluate the whole of image quality without including frame rate. Tearing, frame pacing, MPRT, animation quality etc all matter.

Why does anyone actually care if a PS5 is more or less powerful than a 2060 super anyway?

 

[BRiT]

Can we remain on point from a technical perspective? {{ Please do not answer rhetorical questions }}

 

[Lurkmass]

What is interesting about it?

It's unique in the fact that there's no standardized equivalent or alternatives.

And exactly that was my point (and what I wrote, btw): Did developers who praised the greater possibilities on console Hardware try Vulkan, where there is no one-size-fits all (yes, caveat, there's slight differences) solution but individual extensions.

Most AAA developers prefer D3D12 over Vulkan and there's plenty of reasons from tighter DXGI integration, not having to deal with renderpasses, mutable descriptors, and D3D is also usually the first to standardize new graphics features. Tools like Nsight also work better on D3D12 as well.

Just because vendor specific extensions are comparatively less cumbersome on Vulkan doesn't make for a good reason among developers to use it over D3D12. In, fact IHVs would prefer to extend APIs as little as possible from a maintenance perspective. All titles sponsoring AMD's ray tracing implementation are using D3D12 so AMD decides that it's worthwhile to expose ray tracing hit tokens on that API since ISVs will actually use it. AMD won't do the same for Vulkan because no ISVs currently intend use vendor specific functionality since they didn't ask for it. In theory, vendors can implement any number of extensions they want to expose all hardware functionality on Vulkan. In practice, D3D12 might be arguably closer to matching the capabilities of hardware rather than Vulkan since ISVs will frequently request/use vendor specific functionality and IHVs will implement and maintain the said functionality on the driver side so in the end D3D12 exposes more hardware features when IHVs are more than willing to add extensions to AGS or NVAPI.

I think we should end the discussion here since this is not the appropriate place to speak of these matters ...

 

[DegustatoR]

It's unique in the fact that there's no standardized equivalent or alternatives.

Do we know what it does? The games which use it run on DXR on NV GPUs.

 

[Kaotik]

Do we know what it does? The games which use it run on DXR on NV GPUs.

I'm assuming the only DX12/RT extension they've listed is the one you're talking about, don't know what it does though

Version 6.0 introduces several new shader intrinsics, namely a DX12 ray tracing hit token for RDNA2 hardware for ray tracing optimisation, ReadLaneAt and explicit float conversions.

https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK

 

[OlegSH]

Did some funky measurements today.

Nioh 2 image quality analysis:
https://imgsli.com/NDMxMTM
On the left is a fragment of 9x SSAA "ground truth" image downsampled from 3840x2160 to 1280x720 with bicubic filtering as it doesn't add moire, aliasing and ringing like Lanczos.
On the right is a fragment of 720p DLSS Quality image reconstructed from 853x480 to 1280x720 by accumulating and verifying samples from a large sequence of 853x640 images

Here is another comparison:
https://imgsli.com/NDMxMTQ

Left one is the ground thruth 9x SSAA image again
Right one is native 720p without any kind of AA

And here is the third one just for fun:
https://imgsli.com/NDMxMTA

Left image is 70% rendering resolution (896x504) of 720p + single frame spatial upscale to 720p with Mitchell–Netravali filter + sharpening applied on top of the scaled image aka AMD's Fidelity FX + CAS or whatever it's called
Right image is the DLSS Quality - a sequence of 853x480 images verified by NN

It's kind of obvious which one is closer to the ground truth, but I also decided to make these fragments aligned so that I can capture some image quality metrics to verify that in more mathematical style.
BTW, it was a huge PITA because the character on the screens above was animated and I wanted to catch DLSS in dynamic because on a static image it would have converged to SSAA in a matter of a few frames.

Image fragments with original res and without any additional scaling:
DLSS Quality vs 9x SSAA:
PSNR (more is better): 35.7854
RMSE (less is better): 1064.64
MSE (less is better): 17.2956
MAE (less is better): 651.178

Native 720p No AA vs 9x SSAA
PSNR (more is better): 32.671
RMSE (less is better): 1523.79
MSE (less is better): 35.4302
MAE (less is better): 905.421


The same image fragments, but 9х magnified:
DLSS Quality vs 9x SSAA
PSNR (more is better): 37.2114
RMSE (less is better): 903.449
MSE (less is better): 12.4547
MAE (less is better): 561.239

(Native) No AA vs 9x SSAA
PSNR (more is better): 35.32
RMSE (less is better): 1123.25
MSE (less is better): 19.252
MAE (less is better): 689.93

So DLSS not only looks way more alike 9x SSAA, but it's also closer to SSAA image in terms of image metrics (less math difference between pixels).
Though, no real surprises here since this is exactly how DLSS is supposed to work if everything is set up correctly - converge to the ground truth over time.

I made quite a few screenshots, but, unfortunately, they all were with slightly different camera angles, so I wasn't able to perfectly align them, but in case you are wondering they are all here (including the screenshots above):
https://drive.google.com/drive/folders/1ANx8HURnxFjp…ZYEtf?usp=sharing

 

[DegustatoR]

I'm assuming the only DX12/RT extension they've listed is the one you're talking about, don't know what it does though


https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK

Sounds like a development oriented debug feature more than anything related to actual RT execution. I'm sure that NV has similar features in their NVAPI - in fact wasn't it mentioned around BFV launch that some RT game is using NV's intrinsics?

Anyway, not sure why this is interesting. RDNA2 won't magically become 5X faster in RT through some feature of AGS or AMD would've pushed to include it into DXR 1.1.

 

[Kaotik]

Sounds like a development oriented debug feature more than anything related to actual RT execution. I'm sure that NV has similar features in their NVAPI - in fact wasn't it mentioned around BFV launch that some RT game is using NV's intrinsics?

Anyway, not sure why this is interesting. RDNA2 won't magically become 5X faster in RT through some feature of AGS or AMD would've pushed to include it into DXR 1.1.

Maybe it'll be part of future DXR 1.2, we already know MS allows Xbox devs to skip DXR with its limitations and go straight for the money

 

[pjbliverpool]

Did some funky measurements today.

Nioh 2 image quality analysis:
https://imgsli.com/NDMxMTM
On the left is a fragment of 9x SSAA "ground truth" image downsampled from 3840x2160 to 1280x720 with bicubic filtering as it doesn't add moire, aliasing and ringing like Lanczos.
On the right is a fragment of 720p DLSS Quality image reconstructed from 853x480 to 1280x720 by accumulating and verifying samples from a large sequence of 853x640 images

Here is another comparison:
https://imgsli.com/NDMxMTQ

Left one is the ground thruth 9x SSAA image again
Right one is native 720p without any kind of AA

And here is the third one just for fun:
https://imgsli.com/NDMxMTA

Left image is 70% rendering resolution (896x504) of 720p + single frame spatial upscale to 720p with Mitchell–Netravali filter + sharpening applied on top of the scaled image aka AMD's Fidelity FX + CAS or whatever it's called
Right image is the DLSS Quality - a sequence of 853x480 images verified by NN

It's kind of obvious which one is closer to the ground truth, but I also decided to make these fragments aligned so that I can capture some image quality metrics to verify that in more mathematical style.
BTW, it was a huge PITA because the character on the screens above was animated and I wanted to catch DLSS in dynamic because on a static image it would have converged to SSAA in a matter of a few frames.

Image fragments with original res and without any additional scaling:
DLSS Quality vs 9x SSAA:
PSNR (more is better): 35.7854
RMSE (less is better): 1064.64
MSE (less is better): 17.2956
MAE (less is better): 651.178

Native 720p No AA vs 9x SSAA
PSNR (more is better): 32.671
RMSE (less is better): 1523.79
MSE (less is better): 35.4302
MAE (less is better): 905.421


The same image fragments, but 9х magnified:
DLSS Quality vs 9x SSAA
PSNR (more is better): 37.2114
RMSE (less is better): 903.449
MSE (less is better): 12.4547
MAE (less is better): 561.239

(Native) No AA vs 9x SSAA
PSNR (more is better): 35.32
RMSE (less is better): 1123.25
MSE (less is better): 19.252
MAE (less is better): 689.93

So DLSS not only looks way more alike 9x SSAA, but it's also closer to SSAA image in terms of image metrics (less math difference between pixels).
Though, no real surprises here since this is exactly how DLSS is supposed to work if everything is set up correctly - converge to the ground truth over time.

I made quite a few screenshots, but, unfortunately, they all were with slightly different camera angles, so I wasn't able to perfectly align them, but in case you are wondering they are all here (including the screenshots above):
https://drive.google.com/drive/folders/1ANx8HURnxFjp…ZYEtf?usp=sharing

This is superb work!

 

[OlegSH]

This is superb work!

Thanks!
I spent a few hours trying to capture screenshots with aligned character animation, this was exhausting, next time I will simply capture video. Also, need to write a script for SSIM.

 

[Deleted member 86764]

Did some funky measurements today.

Nioh 2 image quality analysis:
https://imgsli.com/NDMxMTM

On the left is a fragment of 9x SSAA "ground truth" image downsampled from 3840x2160 to 1280x720 with bicubic filtering as it doesn't add moire, aliasing and ringing like Lanczos.
On the right is a fragment of 720p DLSS Quality image reconstructed from 853x480 to 1280x720 by accumulating and verifying samples from a large sequence of 853x640 images

Here is another comparison:
https://imgsli.com/NDMxMTQ

Left one is the ground thruth 9x SSAA image again
Right one is native 720p without any kind of AA

And here is the third one just for fun:
https://imgsli.com/NDMxMTA

Left image is 70% rendering resolution (896x504) of 720p + single frame spatial upscale to 720p with Mitchell–Netravali filter + sharpening applied on top of the scaled image aka AMD's Fidelity FX + CAS or whatever it's called
Right image is the DLSS Quality - a sequence of 853x480 images verified by NN

It's kind of obvious which one is closer to the ground truth, but I also decided to make these fragments aligned so that I can capture some image quality metrics to verify that in more mathematical style.
BTW, it was a huge PITA because the character on the screens above was animated and I wanted to catch DLSS in dynamic because on a static image it would have converged to SSAA in a matter of a few frames.

Image fragments with original res and without any additional scaling:
DLSS Quality vs 9x SSAA:
PSNR (more is better): 35.7854
RMSE (less is better): 1064.64
MSE (less is better): 17.2956
MAE (less is better): 651.178

Native 720p No AA vs 9x SSAA
PSNR (more is better): 32.671
RMSE (less is better): 1523.79
MSE (less is better): 35.4302
MAE (less is better): 905.421


The same image fragments, but 9х magnified:
DLSS Quality vs 9x SSAA
PSNR (more is better): 37.2114
RMSE (less is better): 903.449
MSE (less is better): 12.4547
MAE (less is better): 561.239

(Native) No AA vs 9x SSAA
PSNR (more is better): 35.32
RMSE (less is better): 1123.25
MSE (less is better): 19.252
MAE (less is better): 689.93

So DLSS not only looks way more alike 9x SSAA, but it's also closer to SSAA image in terms of image metrics (less math difference between pixels).
Though, no real surprises here since this is exactly how DLSS is supposed to work if everything is set up correctly - converge to the ground truth over time.

I made quite a few screenshots, but, unfortunately, they all were with slightly different camera angles, so I wasn't able to perfectly align them, but in case you are wondering they are all here (including the screenshots above):
https://drive.google.com/drive/folders/1ANx8HURnxFjp…ZYEtf?usp=sharing

Excellent work. Glad you spent the time capturing a frame of an animated character. I notice a few minor discrepancies, but the DLSS upscale is outstanding. I doubt you'd notice the difference in motion.

I'd be interested to see the results from a rotating camera if it'd be possible to screenshot accurately. It kinda looks like image breakup or blurring on some videos, but figure it's more likely to be YouTube compression.

 

[DegustatoR]

Maybe it'll be part of future DXR 1.2, we already know MS allows Xbox devs to skip DXR with its limitations and go straight for the money

How do you skip DXR and go "straight to the money"? (Don't think that I'm ready to even go into "why".)

 

[Lurkmass]

How do you skip DXR and go "straight to the money"? (Don't think that I'm ready to even go into "why".)

The standard PC DXR pipeline and the Xbox DXR pipeline aren't identical. With Xbox DXR pipeline, you can access to functionality 'beyond' what is offered with the standard PC DXR pipeline. I imagine that's what he meant.

 

[HLJ]

The standard PC DXR pipeline and the Xbox DXR pipeline aren't identical. With Xbox DXR pipeline, you can access to functionality 'beyond' what is offered with the standard PC DXR pipeline. I imagine that's what he meant.

Isn't that mostly "academic" talk, as I doubt any matter of "added functionality" will make up for the performance-deficit in the hardware?

 

[Lurkmass]

Isn't that mostly "academic" talk, as I doubt any matter of "added functionality" will make up for the performance-deficit in the hardware?

Notice how I don't bring up the topic of 'performance' because that's not what the discourse is based on ? The argument was about capabilities while any claims about performance remains ambiguous for the most part ...

 

[HLJ]

Notice how I don't bring up the topic of 'performance' because that's not what the discourse is based on ? The argument was about capabilities while any claims about performance remains ambiguous for the most part ...

Performance is part of the "capabilities" in my world.
You could run DXR on Pascal, but since Pascal lack dedicated hardware...the functionality to run DXR might be there....but the performance is not.
You cannot decouple anything from performance...unless you want a "biased" view shining in a certain way.

So the added "functionality" to me is a moot point, as the performance...well we all know

 

[DegustatoR]

The standard PC DXR pipeline and the Xbox DXR pipeline aren't identical. With Xbox DXR pipeline, you can access to functionality 'beyond' what is offered with the standard PC DXR pipeline. I imagine that's what he meant.

Having additional shader intrisics doesn't mean that the pipelines aren't identical, it means that one of them have additional functions inside the same pipeline. And there can be many reasons for them to be there - not all of them are inherently "interesting".

 

[Lurkmass]

Performance is part of the "capabilities" in my world.
You could run DXR on Pascal, but since Pascal lack dedicated hardware...the functionality to run DXR might be there....but the performance is not.
You cannot decouple anything from performance...unless you want a "biased" view shining in a certain way.

So the added "functionality" to me is a moot point, as the performance...well we all know

Maybe the term 'features' would be a better word since it's absolutely true that console APIs and driver extensions expose more 'features' relative to the standard PC ray tracing pipeline.

Having additional shader intrisics doesn't mean that the pipelines aren't identical, it means that one of them have additional functions inside the same pipeline. And there can be many reasons for them to be there - not all of them are inherently "interesting".

Well they are different pipelines because it's impossible to run the "exact same shaders" on different sets of hardware.