Digital Foundry Article Technical Discussion [2025]

[Charlietus]

Frame gen is a great technology of "secondary" importance.

My problem with it is: higher input lag than with no frame gen. Yes, the difference isn't high, but it's there, so you get to play at 240hz vs 60hz, but technically the responsiveness will be better at 60. And not only that, there is a level of "disconnect" in playing at such high framerates but the expected responsiveness just isn't there.

I also question the need for such high refreshes. I have been passionate about display technology and hardware for a while, but even then sometimes I struggle to see the difference between 60 and 120hz. I know people that struggle to even understand the difference from 30 to 60. This is a niche technology that's valuable to very few people. Still good that it's getting developed.

 

[Boss]

Well no reconstruction tech will ever have 'zero issues', as one of the key metrics to determine the final output is input resolution. DLSS performance mode with an output res of 1080p is something wholly different than DLSS performance mode at an output of 4k, all depends on your criteria.

People can appreciate a feature when it delivers a result that is more cost effective than another method, not so much that it's 'flawless'. I myself have critiqued DLSS in certain games when I first got my 3060, there were definite artifacts that I felt weren't being highlighted, but in general it's gotten quite a bit better - not just from Nvidia's side, but developers as well who are better at looking out for those post-process bugaboos that were more common in earlier games, especially ones that were patched to add DLSS later and it fucks up with motion blur/dof.

Well, sure it might not have zero issues but when we approach a time period where the issues and artifacts aren't easily noticeable, then the glazing can begin. Until then, it's premature and unwarranted. It's fine to appreciate the benefits of DLSS which unfortunately is not what's happening. This hyperbolic praise started when DF claimed that DLSS was better than native and people started parroting it ad nauseum. It wasn't universally true then and it's still not true today. Cleary, there are things DLSS does well but it doesn't do everything well and does many things worse. It's about time to reign in the hyperbole and reclassify that narrative with several addendums.

Oliver does not strike me as someone who often lets his 'emotions get the better of him' in general, but as well he's primarily a Mac and console guy - why would he want to 'glaze' DLSS? He specifically highlights sections where he feels PSSR outperforms DLSS, and you can see how the two can differ greatly from scene to scene. His final conclusion that DLSS is still superior in this title may not entirely match up with the clips provided, but considering that he talked about the issues with both PSSR and DLSS in detail and his history, I see no reason to imply he's being deliberately dishonest. For what purpose?

To quote George from Seinfeld, "It's not a lie if you believe it". I don't think that there's any malicious reason behind his comments. Oliver seems like a stand up dude and I for the most part enjoy his content. I just think it's a case of uttering the DLSS is superior narrative ad nauseum that causes people to arrive at that conclusion even when the data doesn't support it. Also, people weigh issues differently. For some, ghosting is worse than aliasing while for others, the reverse is true.

Nothing can really resolve the issues of non-temporal AA solutions without excessive brute force. TAA was introduced to deal with all the shader aliasing that modern games would have otherwise. So yes, of course DLSS wouldn't exist without TAA, and TAA wouldn't exist if there was actually a performant way to deal with shader aliasing.

And our frames would be free of ghosting and blur. Our textures would be sharp without the need of "artificial sharpening" and 1080p would return to the crispness of old. What a wonderful world that would be.....

 

[Flappy Pannus]

And our frames would be free of ghosting and blur. Our textures would be sharp without the need of "artificial sharpening" and 1080p would return to the crispness of old. What a wonderful world that would be.....

We had it, and with a number of PC releases like most Nixxes ports, we still do. Like I said before, there wasn't a global TAA cabal that blackmailed developers into nearly universally adopting TAA, they adopted it because it was the most performant solution to deal with the increasing problem of shader aliasing from the use of more advanced materials (along with the performance savings of using temporal upscaling techniques for effects the games rely on).

FXAA/SMAA certainly don't cost more than TAA, so developers aren't choosing 'blurry' over 'sharp'. They're choosing potentially less texture detail for the benefit of a massive reduction in specular/shader aliasing across the entire scene that is not possible without TAA/reconstruction/downsampling.

Here's an example of those 1080p 'crisp' halycon days. Uh, no thanks.

 

[techuse]

The DLSS glazing is wild since they both upscalers look bad in this game. While one might be better in certain aspects than others, the end result is that it still looks bad. The problem with all ML upscalers is that for the most part, when fed low input resolutions, they all struggle. This is true regardless of DLSS, PSSR, or XESS version in my experience. Even with DLSS4, it's artifact city based on their own videos and seems rather easy to break. I had to stop keeping track of the artifacts I noticed and it wasn't just their videos either. Other DLSS4 videos from other content producers have similar artifacts.

Alex questioning the purpose of "PSSR"? it's actually very logical. Having your own technology means that you can remain vendor agnostic. If they decided to switch to another vendor for a new console, their games won't have a software dependency to a proprietary upscaler which they'd need to license. I mean, it's pretty obvious that I'm surprised the question was even asked at all. That combined with the fact that AMD still has nothing. They're still trying to beta test fsr4 as we speak....

Finally the comment about not having to discuss DLSS issues on pc is head scratching since DLSS in its current iteration on pc is artifact ridden when fed low input resolutions. Even in their Nvidia preview video with all the DLSS4 glazing, they pointed out noticeable flaws with DLSS3 that were fixed with the new CNN. So to make that comment, quite head scratching indeed....

Based on the footage in that video, I would say DLSS looks a lot better. It is soft, but the PSSR footage is a noisy mess.

 

[Cyan]

hair is of essence to nVidia, now we understand why.

 

[Flappy Pannus]

hair is of essence to nVidia, now we understand why.

Speaking of this, while I definitely notice more light interaction with her hair, there's also quite a bit more breakup with the RTX hair. It's likely something you'd have to see in motion to appreciate, but I'm wondering what the reason is for that.

Anyways, from DF's newsletter - going to be a juicy week:

This week is very, very busy. In fact we will be for a while yet, so let's get on with the updates:

• Rich (ie me) is just about to film the RTX 5090 video with a projected release date of tomorrow on embargo. You'll start to see more of the wealth of data we can now bring into our content, including power, clocks, latency and more. After that, it's directly onto the RTX 5080, I guess, when it turns up.
• John is going to share DLSS 4 impressions with Alex in a video about how triple-A gaming is now accessible to 4K 240Hz and 1440p 400Hz displays. It's all about motion clarity! He's also going to be returning to PlayStation vs Saturn with a look at shooting games next.
• Oliver is working on a range of projects. He'll be taking a look at the PC version of Final Fantasy 7 Rebirth and at some point you'll see a breakdown of Hogwarts Legacy on PS5 Pro. The GRID Legends iOS review currently on early access will go live later this week.
• Alex is taking point on DLSS upgrade coverage with a deep dive into the transformer model, super resolution and ray reconstruction improvements and a look at DLSS 4.
• Tom is taking a bit of a break soon, but not before he's handed in his coverage for Kingdom Come Deliverance 2 on consoles. Alex is keen on looking at the PC version!
• Will is currently being absolutely invaluable in working through our RTX 5090 benchmark data and prepping the Eurogamer review. Our revised benchmark numbers now work in the EG CMS so the bar charts will return!

 

[Boss]

FXAA/SMAA certainly don't cost more than TAA, so developers aren't choosing 'blurry' over 'sharp'. They're choosing potentially less texture detail for the benefit of a massive reduction in specular/shader aliasing across the entire scene that is not possible without TAA/reconstruction/downsampling.

Here's an example of those 1080p 'crisp' halycon days. Uh, no thanks.

Sure sure......



I think I'll have to pass on these.....

 

[DavidGraham]

@Boss
We just explained in the other thread that TAA came into existence because MSAA doesn't work with deferred rendering and because it can't handle shader aliasing that is extensively present with modern materials and modern lighting systems.

DLSS has flaws yes, but they are significantly less than TAA. Ghosting is present in TAA far more than DLSS, aliasing, flickering and blurring are the same. DLSS decreases all of those significantly. DLAA eliminates 95% of them, and DLDSR + DLSS is a step above DLAA, and it preserves sharpness to a large degree.

We know all of this not just from Digital Foundry, but also from other outlets, like TechPowerUp, Hardware Unboxed, ComputerBase, PC Games Hardware, DSOGaming, and many others who do extensive testing and comapring of upscalers, so this isn't Digital Foundry misleading customers at all, it's the whole press coming to the same conclusion.

You can find TechPowerUp articles comparing upscalers for every AAA game release in the last 4 years, and they all point to one single fact: DLSS is better or the same as native.

 

[Cyan]

 

[TopSpoiler]

It would have been nice if they had mentioned whether there was a difference in LOD bias and sharpness filter.

 

[Scott_Arm]

5090 has bigger boost to performance at 4k. Digital Foundry doing native comparisons for RT and raster but also DLSS benchmarks (including latency). Haven't seen other reviews yet.

 

[Scott_Arm]

I do have a criticism about part of this, and maybe I missed the clarification in the video.



This "Native 2160p" result has to be with Reflex off, right? The frame gen results for Reflex on. So are we seeing Reflex off vs Reflex on? I think the native 2160p should have reflex enabled, but I'm not totally clear on the result. I have 90% confidence that this latency comparison is not totally fair.

Edit: I know that DLSS super resolution generally lowers latency a bit, but I think that's mostly from frame time differences, and I wouldn't expect it to be that much less latency than native without reflex.

Edit: I'm wrong. I forgot to think about the upscale step between native and frame gen which would increase base performance and lower latency.

 

[GhostofWar]

It's 35 fps at native 4k vs ~80ish at dlss perf, I wouldn't be surprised if native 4k has reflex on.

 

[Scott_Arm]

It's 35 fps at native 4k vs ~80ish at dlss perf, I wouldn't be surprised if native 4k has reflex on.

You're right. My brain skipped the DLSS perf part. Native to DLSS perf increases performance and lowers latency before the frame gen comes in. I probably wouldn't have blanked on that if they had native, DLSS performance and all of the frame gen metrics on the same visual.

 

[Flappy Pannus]

5090 has bigger boost to performance at 4k. Digital Foundry doing native comparisons for RT and raster but also DLSS benchmarks (including latency). Haven't seen other reviews yet.

BTW if you're wondering why this video is unlisted, it's just youtube:

As I write this, YouTube is still refusing to clear the upload for monetisation - hilarious.

Little disappointing performance diff considering the raw spec increase, particularly with RT heavy titles where I thought it would show more of a gap. I really would that thought these GPU's would be more memory bandwidth bound than they seem to be, albeit I'd like to see some tests with downsampling or titles that give you the option of SSAA just to see if you can really push this aspect.

The DLLS4 latency readings though are quite impressive, there really is barely any difference going from 2X to 4X.

Regardless, never in the class of products for me and most so I just accept this is halo product at halo pricing, at least for gaming - I suspect other uses of the card outside of gaming will show more appreciable differences. What I'm really interested in is the DLSS4 transformer investigation.

 

[trinibwoy]

You're right. My brain skipped the DLSS perf part. Native to DLSS perf increases performance and lowers latency before the frame gen comes in. I probably wouldn't have blanked on that if they had native, DLSS performance and all of the frame gen metrics on the same visual.

It’s still a little disingenuous though as anyone who cares about latency wouldn’t be playing at 35fps in the first place. I think the real point of discussion is whether the increase in base latency from a respectable frame rate to an interpolated frame rate is acceptable. What’s missing from this convo is an absolute latency number that most people would consider OK. It doesn’t matter if framegen adds latency if it remains under that threshold. Somehow we landed on 60fps as an acceptable frame rate. Maybe we can do the same thing for latency.

 

[snc]

I think nobody will use frame gen with dlss perf mode on rtx 5090, surly quality mode is the way to go for this card

 

[IQandHDR]

I think nobody will use frame gen with dlss perf mode on rtx 5090, surly quality mode is the way to go for this card

Yup, never used DLSS Perf mode, I either use DLSS Quality or DLAA if the performance is there.
Same with FSR...
Those settings are irrelevant for a 5090....perhaps a 5070 would see that usage...meh.

 

[Scott_Arm]

If I had a 4k display I'd definitely be using perf mode, but I'd never switch to 4k unless I could get 4k240.

 

[IQandHDR]

If I had a 4k display I'd definitely be using perf mode, but I'd never switch to 4k unless I could get 4k240.

I have a 4K OLED, DLSS Perf has never run on my PC.
Neither has FSR.

DLAA has run a lot.
DLSS Quality has run a lot.
RT too.
HDR even more so.

But DLSS Perf/FSR...nope.

Also what made me turn of the DF 5090 video.
I don't care about 4 x FG DLSS Perf
I do care about 4 x FG DLSS Quality
I would like to know if FG and DLAA can work together.
I do not care about FSR one single bit.

So any review doing DLSS Perf/FSR numbers are not targeting me...an upcoming 5090 buyer
Missed the demographics I'd say.