Support NeoGAF

[PaintTinJr]

It also requires a measurement metric. Compared to the native render as a point of reference, and using the things you listed - higher fidelity resembling supersampled rendering and reduced noise in details - as the metric, the DLSS render is better.

You've chosen something else as the metric, and while I'd agree that for some cases it's important that the result pixel-match the native render, I also struggle to imagine the exact cases where it would be relevant considering the supposed deviation is in terms of pixels, on a 4K image.

And I also can't quite find what you're referring to in regards to examples, can you point me to the specific comparison? Seeing as a lot of comparisons are from normal gameplay, and recording both DLSS on and DLSS off footage is functionally impossible, it's not out of the question for the comparison shots to be slightly different in timing and scene composition.

If they haven't bothered to ensure that both images are of identical runs and game states, then the comparison is worthless, and shows they are more concerned about perception of better than actual doing the leg work to level the playing field to show the algorithm can handle both low frequency inference perfectly, along with improving high frequency detail at the same time,

For signal processing, low frequency detail is the base signal(and more important), and is the more important metric when it has errors IMO - because images are sampled and built from low to high in an additive way, with the highest details - that can't be perceived - always being left unsampled.

The speaker shows that none of the results are perfect IMO - with the DLSS biasing towards fixing the high frequency at the detriment of losing low frequency by the mesh occluding that low frequency detail, and the other two giving both high and low frequency detail equal importance, but even at native, the resolution is too low to sample the speaker mesh correctly, and leaves both meshes looking like a rendering artefact.

 

[01011001]

This is crazy. Somebody needs to explain how DLSS works, how can it look better than native res?

it uses multiple samples and is trained with high resolution images that it then uses to complete missing details.

the algorithm knows that the bridge needs to have cables going down there so it draws cables where they belong.

 

[Fafalada]

The model being generalized is why it can solve things like ghosting. Because ghosting is a repeating pattern that the AI can learn to recognize and eliminate during reconstruction

That's not really it though - ghosting is a result of reprojection errors - the source of which varies (transparent particles were an obvious one, but not even the most egregious in my experience with DLSS).

 

[Md Ray]

 

[Whitecrow]

Every comparison is a clickbait. FSR doesnt hold a candle against DLSS in terms of quality.

 

[MightySquirrel]

DLSS really is some sort of MAGIC .


TAA native vs DLSS Perf. [DLSS +50% higher frame rate] :

BONUS Native TAA vs DLSS perf. :

 

[Md Ray]

DLSS really is some sort of MAGIC .


TAA native vs DLSS Perf. [DLSS +50% higher frame rate] :

BONUS Native TAA vs DLSS perf. :

What resolution?

How does it look in motion?

 

[Low Moral Fibre]

That Hardware Unboxed video really undersells how bad FSR and TAA looks in Avengers although that is probably an issue with the in-game TAA rather than FSR.

I don't know what AA Anno 1800 uses but I can't tell the difference with FSR on unless I pixel peep and there is a really nice FPS boost.

I guess the problem with FSR is that most games currently seem to use TAA so if the implementation is poor then FSR just exacerbates that while much of the better than native talk with DLSS is because it sidesteps the potentially crappy TAA.

Anyhow I'm pleasantly surprised with FSR as I went back to playing Monster Hunter which I think is DLSS 1.0 and it is an absolute mess with DLSS on with artifacting everywhere. That FSR is even remotely comparable is great even if it is clearly behind DLSS 2.x at this stage.

 

[MightySquirrel]

What resolution?

How does it look in motion?

5K. Second comparison was a joke. I think the card just chocked on bandwidth at that res, as it was running at ~45 before I enabled DLS and when went back to native saw 26 instead of 45, was pretty confused. Still ~50% uplift at high resolution is insane when IQ looks nearly identical.

And yeah I'm surprised nobody talks about how much better DLSS handles motion over TAA, at least in RDR2. Amybe it's just that particular game?

Here's an example where I pressed the right stick all the way to the left and took a shot at full rotation speed: Notice how much cleaner it looks with DLSS!

TAA vs DLSS full speed camera pan. DLSS looks so clean in comp. You can easily tell by simply playing that TAA native is blurry.

 

[Md Ray]

5K. Second comparison was a joke. I think the card just chocked on bandwidth at that res, as it was running at ~45 before I enabled DLS and when went back to native saw 26 instead of 45, was pretty confused. Still ~50% uplift at high resolution is insane when IQ looks nearly identical.

And yeah I'm surprised nobody talks about how much better DLSS handles motion over TAA, at least in RDR2. Amybe it's just that particular game?

Here's an example where I pressed the right stick all the way to the left and took a shot at full rotation speed: Notice how much cleaner it looks with DLSS!

TAA vs DLSS full speed camera pan. DLSS looks so clean in comp. You can easily tell by simply playing that TAA native is blurry.

Oh, I immediately picked up on the sharpness that DLSS brought over the native image's TAA blurriness in my tests too.

But outside of quality mode, especially in performance some of the elements like the tree branches had a lot of noticeable shimmering in motion, so I was wanted to know how it was for you. I did my tests in 4K, might give 5K a try as well.

Is your GPU 3080?

 

[Evilms]

 

[Md Ray]

Wow,

540p = A STARK FAMILY HISTORY
720p = ᓍ SWUHNVKFSDL ɥ!sʇoRY

 

[MightySquirrel]

Oh, I immediately picked up on the sharpness that DLSS brought over the native image's TAA blurriness in my tests too.

But outside of quality mode, especially in performance some of the elements like the tree branches had a lot of noticeable shimmering in motion, so I was wanted to know how it was for you. I did my tests in 4K, might give 5K a try as well.

Is your GPU 3080?

Oh yeah I noticed at 2560x1440 even with dlss quality setting that there was some shimmer on far trees, but at 5K + DLSS perf setting it looks insane. Like looking at supersampled image.

 

[Evilms]

a little off topic

Nvidia brings RTX and DLSS technology to MediaTek ARM processors with Wolfenstein demo at GDC 2021, very impressive

 

[Hypno285]

Why is native looking like shit?

 

[Md Ray]

Oh yeah I noticed at 2560x1440 even with dlss quality setting that there was some shimmer on far trees, but at 5K + DLSS perf setting it looks insane. Like looking at supersampled image.

Shit, didn't think 5K would be possible on a 3070 @ 60fps.

 

[Armorian]

Why is native looking like shit?

DLSS and native TAA are handling transparencies completely different (native = not at all). FSR is just worse version of native so...

 

[Md Ray]

a little off topic

Nvidia brings RTX and DLSS technology to MediaTek ARM processors with Wolfenstein demo at GDC 2021, very impressive

Nice, an RTX 3060 driven by an ARM octa-core (MediaTek) CPU to accelerate ray tracing.

 

[Armorian]

Native

DLSS Q

Native

DLSS Q

Transparencies handled differently again ^

DLSS is IMO oversharpened in this game

 

[ratburger]

DLSS destroys a lot more details than FSR & Native, but the fanboy brigade is never eager to point those out, instead only showing wires. Don't forget that FSR is not a form of AA unlike DLSS, which means you can pick and choose your poison. Like I said, there's pros & cons to all these methods.

That could be a LOD bias issue, wherein the game doesn't realize that the image is going to be upscaled by DLSS so uses lower-res textures. There are a few games, like Cyberpunk, that have that problem. It can be fixed by using Nvidia Inspector and setting the override LOD bias to -3 (for Performance, probably -1 for Quality, but not sure).

 

[JackMcGunns]

Wow DLSS is just bitch slapping FSR!

And also Native 4K, so it's not even fair LMAO

 

[Roman Empire]

DLSS is on another level and the tech still has a long way into improving. FSR is still in its infancy but the game between the both and even compared to the default resolution is just too great. DLSS all the way, at least for now. The more options, the better.

 

[Stitch]

DLSS destroys a lot more details than FSR & Native, but the fanboy brigade is never eager to point those out, instead only showing wires. Don't forget that FSR is not a form of AA unlike DLSS, which means you can pick and choose your poison. Like I said, there's pros & cons to all these methods.

It looks like Native and FSR can't render the mesh correctly, like they can't correctly render fences and the bridge wires and the DLSS one is actually the correct one..

 

[sendit]

DLSS is on another level and the tech still has a long way into improving. FSR is still in its infancy but the game between the both and even compared to the default resolution is just too great. DLSS all the way, at least for now. The more options, the better.

Not surprised. AMD also gets bitched slapped in Ray Tracing performance.

 

[Kupfer]

Those speakers look like hot trash native or fsr and actually look like a speaker grill in dlss. The dlss is way way way above in that speaker pic like it's no contest at all. Native and fsr the grill is just sparse and hot trash.

You missed the crucial point. It's not about making the speaker look like you imagine a speaker to look like, even if the DLSS speaker looks good on its own, ALL the details that are behind the grille are lost. That the DLSS speaker still looks good is nice, but I don't want details and whole geometry to be swallowed up.

It looks like Native and FSR can't render the mesh correctly, like they can't correctly render fences and the bridge wires and the DLSS one is actually the correct one..

It all depends on how the loudspeaker is supposed to look on the part of the developers. The fact that you can clearly see behind the grille and recognize details distorts DLSS's image.
Real life for reference :

 

[Roman Empire]

Not surprised. AMD also gets bitched slapped in Ray Tracing performance.

Yeah, that is true too. But I like the idea of AMD aspiring to "compete" with Nvidia in this field. Nothing better than a competition to bring forth new ideas and techniques which will optimize the technology even further.

 

[Sean Mirrsen]

You missed the crucial point. It's not about making the speaker look like you imagine a speaker to look like, even if the DLSS speaker looks good on its own, ALL the details that are behind the grille are lost. That the DLSS speaker still looks good is nice, but I don't want details and whole geometry to be swallowed up.

It all depends on how the loudspeaker is supposed to look on the part of the developers. The fact that you can clearly see behind the grille and recognize details distorts DLSS's image.
Real life for reference :

It's possible that it's a texture transparency issue. I notice how in the DLSS shot it seems that the grille texture has an opaque black background, when it should be transparent.

It wouldn't be the first time texture transparency was an issue in a DLSS implementation.

 

[Werewolfgrandma]

You missed the crucial point. It's not about making the speaker look like you imagine a speaker to look like, even if the DLSS speaker looks good on its own, ALL the details that are behind the grille are lost. That the DLSS speaker still looks good is nice, but I don't want details and whole geometry to be swallowed up.

It all depends on how the loudspeaker is supposed to look on the part of the developers. The fact that you can clearly see behind the grille and recognize details distorts DLSS's image.
Real life for reference :

I actually changed my mind about that yesterday. I do believe it looks better, but not correct.

 

[StateofMajora]

People acting like dlss is perfect or better than native are on something. It's important first of all to know what you're comparing dlss to.. if it's a flawed TAA solution that scrubs away detail that's one thing, but a pristine native solution with smaa 1x or msaa would not have detail scrubbed away nor any temporal artifacts.

DLSS is very impressive and is quickly evolving into the ultimate upscaling technique, but it's not there yet. However, it's clear that it adds a lot of quality to sub native resolutions and is pretty much the best at doing that (the only contender as I see it being Insomniac temporal injection)

 

[FukuDaruma]

Damn, that horse texture looks lowres af. Are these top quality Ultra settings? RDR2 horses are supposed to be the best horses in any videogame...

 

[PaintTinJr]

Giving the Speaker Separates picture comparison some more thought...I'm wondering if what that really shows is that DLSS 2.2's frame-rate and high frequency detail enhancement benefits are going to be short lived - just while FSR is faithfully trying to reproduce 4K native of last-gen AAA visuals, where IQ at 4 or 8K can get a guess enhancement that typically looks nicer than the native image.

Based on the way DLSS handles the speaker mesh at the comparison draw distance, I would expect most of the frame-rate boost of DLSS to be tied to the mesh (wrongfully) occluding most of the speaker separate's internals. But had the camera moved closer to the Speaker separate, the frame-rate gains of DLSS would drop, as the high frequency mesh becomes comparative low frequency signal compared to the internals, and the DLSS image would have to render everything, or start to look vastly inferior to the native and FSR images - because the high frequency mesh artefacts at the previous viewpoint would (instead) render perfectly, when the mesh became a comparatively low frequency signal at a closer viewpoint.

Equally, moving the viewpoint away from the speakers we should AFAIK expect the mesh to occlude more of the internals in all images; but as the DLSS image is already (wrongly) occluding virtually all of the internals as is, then the internals would be fully occluded from a more distant viewpoint.

By comparison the native - and subsequent near identical FSR image - would show more occlusion from the mesh making the overall composition perfect - but probably introducing heavy mesh aliasing artefacts because of the last-gen native IQ or need for 8K native rendering for the mesh, or bit of both.

So if those scenarios would be as I now expect, then when UE5 (nanite/lumen +hw RT) IQ is the normal for games, and the reference training image for comparison is 8K native, then I would expect FSR to continue to reproduce the native image, where as I would expect DLSS to be unable to improve the native IQ and struggle to faithfully reproduce the native image - and in the event it could, the performance gains would probably be taken via the increased fill-rate from the increased number of pixels being inferred in the DLSS output.

 

[SantaC]

Atleast FSR works on nvidia cards also.

Nvidia is too cheap to try and get DLSS to amd cards.

 

[Amiga]

symbolic photo finish win for DLSS but a loss for Nvidia. FSR works in emulating native rendering. it works from v1.0, and at this rate will be standard more than DLSS. this is all that matters for normal gamers. Less justification for Nvidia premium prices.

now if we compare equivalent GPUs with the upscaling things will get interesting. 3070DLSS vs 6800xtFSR.

Nvidia just has the RT advantage for now and could lose big next time as AMDis expected to offer better RT acceleration and improve FSR.

 

[Armorian]

symbolic photo finish win for DLSS but a loss for Nvidia. FSR works in emulating native rendering. it works from v1.0, and at this rate will be standard more than DLSS. this is all that matters for normal gamers. Less justification for Nvidia premium prices.

now if we compare equivalent GPUs with the upscaling things will get interesting. 3070DLSS vs 6800xtFSR.

Nvidia just has the RT advantage for now and could lose big next time as AMDis expected to offer better RT acceleration and improve FSR.

Improved FSR will be as hard to implement in games as DLSS, so advantage of adding it in few minutes will be gone.

Premium prices? Aside 3090 ridiculous price both AMD and Nvidia MSRP prices are very similar for they performance. AMD also asks a lot for their CPUs so they are not as "good company" as they used to be.

 

[Amiga]

Premium prices? Aside 3090 ridiculous price both AMD and Nvidia MSRP prices are very similar for they performance. AMD also asks a lot for their CPUs so they are not as "good company" as they used to be.

the buyers in the market decide the price not Nvidia or AMD. no business will give away stuff. Nvidia for long had the "premium" reputation so AMD had to go for price sensitive buyers. the less advantages Nvidia have the less price premium they can command. this is what the real fight is about between the 2. Nvidia are not complacent like Intel though, I'm sure they will react hard. but I don't think it will be enough to maintain the quality gap.

 

[alucard0712_rus]

symbolic photo finish win for DLSS but a loss for Nvidia. FSR works in emulating native rendering. it works from v1.0, and at this rate will be standard more than DLSS. this is all that matters for normal gamers. Less justification for Nvidia premium prices.

now if we compare equivalent GPUs with the upscaling things will get interesting. 3070DLSS vs 6800xtFSR.

Nvidia just has the RT advantage for now and could lose big next time as AMDis expected to offer better RT acceleration and improve FSR.

Are you serious?

 

[Amiga]

Are you serious?

considering you didn't argue against what I said I think I'm on the right track.

 

[Omega Supreme Holopsicon]

FSR is just worse in every way. I'd rather just play with a lower native resolution than try and upscale, if those are the results. DLSS 2.2 is just other-worldly good at this point.

I think the only thing that could realistically compete with DLSS would be a significant evolution or update of fractal algorithm image upscaling.

The universe is made of information, all is information, optimal physics of information transmission and processing often results in fractal like structure following certain laws, as observed in vascular system and other naturally occurring optimal structures. Even the human brain operates near fractality according to some, in fact nondeterministic fractals might potentially evade classification as fractals due to they appearing as nonfractal due to the minute detail differences(but this is a hypothesis).

 

[PaintTinJr]

Improved FSR will be as hard to implement in games as DLSS, so advantage of adding it in few minutes will be gone.

Premium prices? Aside 3090 ridiculous price both AMD and Nvidia MSRP prices are very similar for they performance. AMD also asks a lot for their CPUs so they are not as "good company" as they used to be.

Why would a technique that gets very close to native - not just scale up with IQ and resolution - need to improve?

The faults with FSR currently appear to be that it doesn't guess at trying to improve IQ of the original native image - and maybe room to improve marginally in matching native at lower performance/memory cost. As we move into next-gen graphics and higher native resolutions, FSR should be both a better image compared to native, and visually look better than DLSS.

 

[Sean Mirrsen]

Why would a technique that gets very close to native - not just scale up with IQ and resolution - need to improve?

The faults with FSR currently appear to be that it doesn't guess at trying to improve IQ of the original native image - and maybe room to improve marginally in matching native at lower performance/memory cost. As we move into next-gen graphics and higher native resolutions, FSR should be both a better image compared to native, and visually look better than DLSS.

It will be hard for FSR to look better than the native image, considering it has no data to work from besides the native image.

At the same time, DLSS achieving better visuals for the same performance, simply means there's room for achieving better performance, or using weaker hardware (or lower clocks on the same hardware) to achieve the same visuals and performance. Which will be rather important for any mobile gaming hardware Nvidia ends up making - a point made extra interesting now that they've showcased that DLSS and RTX both work with ARM.

 

[Omega Supreme Holopsicon]

It will be hard for FSR to look better than the native image, considering it has no data to work from besides the native image.

At the same time, DLSS achieving better visuals for the same performance, simply means there's room for achieving better performance, or using weaker hardware (or lower clocks on the same hardware) to achieve the same visuals and performance. Which will be rather important for any mobile gaming hardware Nvidia ends up making - a point made extra interesting now that they've showcased that DLSS and RTX both work with ARM.

DLSS can potentially work in theory for brand new games without necessarily being trained on these new games. So the NN is approaching some mathematical structure that exists and embodies some sort of ideal upsampling that can work on even arbitrary input outside training data. The question is what is the nature of that structure? Are there alternate algorithms that can make use of this data or mathematical structure as well?

We can imagine the case of infinite training with the finite 4k or 16k hdr image set, after infinite training with all possible images, a finite NN structure will emerge embodying perfect upscaling for all possible resolutions. What is its connectivity like? Is there a simpler order, a fractal like formula that can embody this infinite amount of information within a finite body akin to how nanite handles virtually infinite geometry within finite amount of pixels?

 

[PaintTinJr]

It will be hard for FSR to look better than the native image, considering it has no data to work from besides the native image.

At the same time, DLSS achieving better visuals for the same performance, simply means there's room for achieving better performance, or using weaker hardware (or lower clocks on the same hardware) to achieve the same visuals and performance. Which will be rather important for any mobile gaming hardware Nvidia ends up making - a point made extra interesting now that they've showcased that DLSS and RTX both work with ARM.

I'm assuming It will be impossible for DLSS to look better than native - and by extension the FSR reproduction - because as IQ and resolution improves, the areas in which DLSS can make a guess at improving IQ will get very small compared to the wrong inferences it makes - like in the speaker picture comparison.

On mobile images on small screens with retina level or better displays exist - minification naturally removes most of the areas in which DLSS currently guesses better than native, so I don't expect that to be a win area for them either.

Guessing and making something better is nice at the moment, but game rendering IMHO should be deterministic, as an incorrect inference impairs a gamer's ability to read the action and play accordingly. Where it might be acceptable for a dog detection app using ML to have a ~95% success rate, such an error rate for rendering 100 goons with their backs turned, in a game like Batman - and getting it wrong 5 times, before they turned around - wouldn't be acceptable, no matter how nice it improved the flowers at reception in the psych ward compared to FSR IMO.

For me, the SteamDeck has put the cat amongst the pigeons by using AMD x64 in a handheld - with Apple and nvidia going all in for ARM. With phones overlapping with handheld gaming, I suspect someone will go with a Zen3 chip in a smartphone soon - if the Deck resonates with gamers and demonstrates comparable or better performance per watt - and we could eventually see ARM replaced by x64 to reignite the smartphone/tablet market as they continue to overlap laptop/desktop computing functionality, so DLSS being tied to Nvidia makes its future even less certain IMO.

 

[MrHydra933BR]

Obviously no one should compare an AI algorithm against a single-frame one.

 

[Sean Mirrsen]

I'm assuming It will be impossible for DLSS to look better than native - and by extension the FSR reproduction - because as IQ and resolution improves, the areas in which DLSS can make a guess at improving IQ will get very small compared to the wrong inferences it makes - like in the speaker picture comparison.

Conversely, I am of the opinion that even if the AI deviates from "truth", it's better if it ends up presenting a more coherent image at a good framerate. And as the tech improves, there will be fewer and fewer areas where the AI 'guess' is measurably wrong, especially as the resolution increases and the errors at pixel and sub-pixel level become less and less relevant.

To me the signature ability of this method, is reconstructing usable, passable high-resolution imagery, from downright absurdly low base render resolution. If Nvidia comes up with an SoC that approximates the one in the Steam Deck, but also comes equipped with Tensor cores, the hypothetical Deck-like device (we may as well colloquialize them as "decks", these hybrid tablet PCs with specialized hardware) will have a tremendous performance ceiling when it comes to scalability. However ill-equipped its GPU would be to handle 'docked' output to higher resolutions, it won't matter because DLSS could pick up arbitrary amounts of slack, upscaling to 4K from 720p or 540p if need be, with appropriate reduction in visual fidelity but maintaining performance, which is the important thing.

Put simply, there are fewer limits to the range and ability of DLSS, than there are to FSR. FSR can only go so far - it is limited by being an upscaling algorithm working off of just rendered imagery. A DLSS-enabled GPU will take far longer to "age", as it were, as it would be able to arbitrarily scale down its render resolution and still come out with adequate image quality at great performance on the output - especially as the tech improves. A fact extremely important for any mobile GPU solution.

 

[Lethal01]

I think the only thing that could realistically compete with DLSS would be a significant evolution or update of fractal algorithm image upscaling.

The universe is made of information, all is information, optimal physics of information transmission and processing often results in fractal like structure following certain laws, as observed in vascular system and other naturally occurring optimal structures. Even the human brain operates near fractality according to some, in fact nondeterministic fractals might potentially evade classification as fractals due to they appearing as nonfractal due to the minute detail differences(but this is a hypothesis).

How high are you?

 

[Omega Supreme Holopsicon]

How high are you?

I do nootropic cocktails not psychedelics

Iirc the young human brain can consume over 40% of the body's energy vs 20% in the adult phase. Most adult brains operate at a small fraction of the full abilities of the human brain.

 

[PaintTinJr]

Conversely, I am of the opinion that even if the AI deviates from "truth", it's better if it ends up presenting a more coherent image at a good framerate. And as the tech improves, there will be fewer and fewer areas where the AI 'guess' is measurably wrong, especially as the resolution increases and the errors at pixel and sub-pixel level become less and less relevant.

I'm not expecting that to happen given that world space scene complexity is going to increase, and if DLSS 2.2 can't handle partial occlusion of a speaker mesh, then for instance, what likelihood is it going to correctly infer multi bounced RT lighting behind a semi transparent piece of glass or ice?
Wrong inference would be my very definition of an incoherent image - in respect of a game rendered image

To me the signature ability of this method, is reconstructing usable, passable high-resolution imagery, from downright absurdly low base render resolution. If Nvidia comes up with an SoC that approximates the one in the Steam Deck, but also comes equipped with Tensor cores, the hypothetical Deck-like device (we may as well colloquialize them as "decks", these hybrid tablet PCs with specialized hardware) will have a tremendous performance ceiling when it comes to scalability. However ill-equipped its GPU would be to handle 'docked' output to higher resolutions, it won't matter because DLSS could pick up arbitrary amounts of slack, upscaling to 4K from 720p or 540p if need be, with appropriate reduction in visual fidelity but maintaining performance, which is the important thing.

Won't that issue be moot - by virtue of UE5's features being limited to high-end x64 PC and Ps5/XsX? And so the scaling would be of last-gen/cross-gen visuals like RDR2, and any future SoC by Nvidia will surely be ARM, which won't run x64 code steam library, like the SteamDeck.

Put simply, there are fewer limits to the range and ability of DLSS, than there are to FSR. FSR can only go so far - it is limited by being an upscaling algorithm working off of just rendered imagery. A DLSS-enabled GPU will take far longer to "age", as it were, as it would be able to arbitrarily scale down its render resolution and still come out with adequate image quality at great performance on the output - especially as the tech improves. A fact extremely important for any mobile GPU solution.

Based on the UE5 target performance, both DLSS and FSR will working to enhance 1080p/1440p native renders to 4K, for now, and probably 4K to 8K after that IMO.

With native quality that high feeding into FSR and DLSS, the problems DLSS currently has like handling the complexity of the speaker will be too big to compete IMO - in the context of the increased IQ from UE5 nanite's ability to constantly render triangles at or around 1 per pixel native, so any DLSS errors will be bigger than the native errors, and the inference enhancements will be similarly small to see a diminishing return is what I would expect.

Although, hopefully you are correct as I recently paid over the odds for a 12GB RTX 3060, as a hold over until I complete a workstation rebuild, so would prefer if it does age well.

 

[FearlessBelgian]

Dlss is so much better...

 

[Sean Mirrsen]

Based on the UE5 target performance, both DLSS and FSR will working to enhance 1080p/1440p native renders to 4K, for now, and probably 4K to 8K after that IMO.

With native quality that high feeding into FSR and DLSS, the problems DLSS currently has like handling the complexity of the speaker will be too big to compete IMO - in the context of the increased IQ from UE5 nanite's ability to constantly render triangles at or around 1 per pixel native, so any DLSS errors will be bigger than the native errors, and the inference enhancements will be similarly small to see a diminishing return is what I would expect.

I really don't see how UE5 factors into things, considering it's far from the only engine around. Unity games will still be made, UE4 games will still be made, things like Unigine or Godot may well catch up and earn some market space.

And you seem to be basing a little too much of your viewpoint on the speaker mesh thing. If you look closely at it, you'll clearly see that the mesh is rendered opaque rather than transparent. It seems like it can be an implementation error, rather than a reconstruction one, especially since the occasional texture quality and opacity bug is one of the few remaining known issues with DLSS.

DLSS won't have issues with sub-pixel triangles either, I don't think - even at 8K, because it trains on 16K. Its only major limit is the reliance on Tensor computing, or rather Nvidia's reluctance to try and decouple it from that reliance. Who knows, maybe we'll eventually see DLSS go the way of PhysX, able to work anywhere but doing it better on dedicated hardware - hardware that Nvidia is currently quite well positioned to make. And with lower-power systems continuing to exist, like the Switch, the Deck(s), and the mobile platforms, there will always be a place for a dedicated solution that can extract high-detail imagery from low-resolution rendering, and do so dynamically regardless of input or output size, adjusting on the fly if needed.