Welcome to the Vergecast, the flagship podcast of deep learning super sampling. I'm David Imel. Nvidia's DLSS5 technology was released on September 3rd, and it's notably different from the ways it's been doing super sampling in the past. Super sampling is a technology used to boost video game fidelity and performance on existing hardware. There's been a lot of pushback from some people in the gaming community and even the developer community about this. So, I wanted to bring on Linus Sebastian and David Gauthier from Linus Tech Tips to talk to me about what DLSS5 is and why pretty much everybody is so up in arms about it. But first, here's what's happening on The Verge today. This is 90 seconds on the Verge for Tuesday, September 22nd, 2026.
Beats have announced the new Beats 360 headphones, which are specifically designed for working out. They come with IPX4 water resistance, removable ear cups, and replaceable batteries. Senior reviewer Victoria Song says they can get a bit icky after a 40-minute workout, but the ear cups are pretty easy to clean once they get too gross. They don't come cheap though. At $350, they're competing with headphones from Sony and the original Sonos Ace. Elsewhere in the exercise world, Peloton have announced a cheaper foldable treadmill called the Tread Flex. It's still pretty expensive at $2,195, but this time it can save 50% of the floor space when it's folded up. Peloton have also announced Whoop integration. So if you're a Loyal Fitness Band subscriber, you can get much more accurate treadmill readings.
Next, Microsoft has announced 268 layoffs as part of its 3,200 layoffs it has planned for this year. As part of the reorg, Halo development is moving under Activision and Obsidian is moving under Bethesda to focus on the Fallout games. Not even Minesweeper and Solitaire are safe. Those games are moving to Candy Crush developer King. And finally, Apple is opening a live music venue in London next week called the Apple Music Hall. It was built under the company's European headquarters and has room for 600 audience members. But if you can't make it to London, I'm sure you're going to be able to visit this thing in the Vision Pro pretty soon. A flight to London would still probably be cheaper, though. You can read more at theverge.com. That's 90 seconds on the Verge for Tuesday, September 22nd, 2026.
All right, Linus Sebastian, David Gauthier, welcome to the Vergecast.
Yeah, I'm super glad to have you guys here. It's been a while since I've seen you, so it's nice to see you first of all, but we really wanted to bring you in because there's this new thing that just dropped, Nvidia DLSS5, super sampling sort of, right? Yeah, we've lost track a little bit of what the SS in DLSS stands for. I think that's pretty clear.
Yeah, we've had DLSS before, which was sort of a way to upscale content when you had a lower-tiered GPU. Can you explain what DLSS 5 is and why it's different from what we've had before?
Yeah. Well, DLSS in the past meant we've got this lower resolution that we're rendering at internally and then we're using, you know, some form of machine learning to take that and extrapolate. You know, kind of like way back in the day when you would have like a low-resolution image that you took on your one-megapixel digital camera and you'd kind of try to stretch it out and then, you know, sharpen it and fill in the missing information. It was like that but powered by machine learning. So towards the end of it with DLSS 4.5, actually, getting pretty good.
DLSS5 takes that and instead of just, you know, adding pixels that weren't there, it seems to be adding elements of the scene that weren't there or adding colors that weren't there. And Nvidia has strongly hinted, not just hinted, they have said outright that this is where they believe the future of rendering is going because they just basically have been saying that Moore's law, where we can just have more and more transistors and bigger and bigger chips that can traditionally render better and better every year, is completely dead. That's Nvidia's kind of stance on that. This concept of neural rendering is the future where we take a relatively low-fidelity scene and then we use our best guesses based on a trained AI model as to what exactly the intent was to have there.
One of the demos of this technology that really stood out to me, and this was years ago, they did this demo where they had this almost MS Paint-like drawing application. Do you guys remember this?
Oh yes. And they drew this like mountain scene and then just like a blue line squiggle through it and then they just press the button and it goes, 'Wow, beautiful mountains and a river,' right?
And that seems to have, for longer than I think most people outside the company have realized, that seems to have been Nvidia's vision. So I, as a, and it seems like earnestly they believe that this will be a benefit to developers and gamers. So, as a developer, I could just say barrel and then a beautifully rendered barrel would just get crapped out onto my screen. That looks photorealistic essentially because I don't actually have to render all those pixels and I can do it so cheaply.
Now, David, our David, David Gauthier, actually spent a lot more time with DLSS5 than I did on games that Nvidia didn't authorize. And it seems like at least at this stage we're falling well short of, you know, magic barrel and oh, here's a leather jacket and it just suddenly is trained on what a leather jacket is supposed to look like and is miraculously photorealistic. So, I think it's a little disingenuous for people to say, well, it's just an Instagram filter. But, you know, David, you looked at it across a much wider range of games than I did. Where does it sit on that spectrum of magic neural rendering to Instagram filter in your mind?
I think its biggest limitation is the fact that it has to be rendered 60 times at least a second. And so that tiny little window to be able to generate the uplift of AI really limits the amount of, you know, creativity that the AI can do. Even Nvidia is trying to distance it in terms of messaging from being generative AI, but as it is right now, they've trained the program to be able to separate objects and you can control each object individually and you can apply different amounts of the, you know, quote-unquote filter to those objects and it doesn't transform them in the sense that it can take a crude drawing of a barrel and make it work. It needs a pretty good input. They said the better input in, the better output out. And I noticed that in my testing that if I gave it higher-fidelity graphics, it looks significantly better after the AI. That being said, it still has that AI look. And if you have that immediate disgust reaction to AI-generated content, this won't change your mind about it.
I have so many questions here because you mentioned you've tried this on a bunch of different games. I watched your guys' video that you put out where you tested on Expedition 33, you tested on a Final Fantasy game and it did seem to upscale the textures to an extent or maybe it was like a dehaze filter a little bit. You know, the haze kind of went away, the contrast went up. Does that not take a lot of creative power away from the artists and the developers and the way that they want these characters to look because to your point in your video, you showed one character that was supposed to be kind of an anime-style character and they ended up just looking like a real person. So where do you think the line is there right now?
You know, the conversation around artist intent is really complicated because the DLSS that we were using were leaked. Only NBA 2K27 has actually been tuned with all the different sliders that Nvidia makes available to developers to decide how they want to apply these effects. So, every other game that we looked at was a combination of our best guesses for what might look pretty okay and what whoops definitely doesn't and also us kind of goofing around with them to see what our range of best-case scenario and worst-case scenario might actually look like here.
Mhm. I think it's going to be a really contentious issue even with game developers. I think some of them see it as the AI slop filter. You turn it on, everything sucks. And then others are coming to the defense. The creative director of Kingdom Come Deliverance 2 went on Twitter or X to defend it and said like this is what we wanted our game to look like. We just didn't have the technology or the resources to make it happen.
Well, they made that a lot worse. So, uh, don't get us started on Star Wars. Don't do it.
But I think right now and I think in the immediate future, this will only be good in games that are chasing full photorealism and anything where there's a stylized look doesn't work. Even games that are kind of halfway there, like Expedition 33, where it's realistic-proportioned characters, but there's sort of a dreamlike quality to it. It sort of takes away from that. And another thing I noticed is that whatever AI they're or whatever data set they're training it on is sort of sucking the color out of it in the same way that lazy color correction in film is doing. And so I think they're using a kind of generalized data set that is like, oh, that's what film is doing. It's trying to make it look more like an actual movie without doing the detail work to make it actually pop.
And I just want to jump in and say I don't want my games to look like movies.
Right. Yeah. Games should look like games and movies should look like movies except when the movie is, you know, the Super Mario World movie or when a game is really like committed to, you know, that gritty realistic filmic look and it's a very, you know, cinematic art style and very story-driven, like then okay, yeah, fine, you know, make my game look like a movie. But these are two different mediums and we shouldn't just go, okay, well, yeah, they're the same now, right?
Yeah. Now, since this is only an Nvidia technology, are we going to enter a world where AMD gamers have a different graphical experience than Nvidia gamers?
Okay. Now, you're getting into like the big questions and okay. Okay. Immediately, no. Because the performance hit of this is already pretty catastrophic on a 5090. And I made an off-hand remark, I think, in the video that it was going to be, I think I said five years or something like that before anyone's like really using this. And someone responded to me. They go, 'What are you talking about? I'm using it today and I love it. I have a 5090.' And I'm like, 'How many people have a 5090?' Right. Yeah. Exactly. You know, because this is not a real mainstream technology until, you know, the last generation's 60-series card can run it in some kind of meaningful way.
And that will not happen until we are a couple launches in and Nvidia launches a new GPU series like every two years now. So that was kind of what I was trying to say there is that right now, you know, this is very similar to, I would say, the early days of the more traditional DLSS that was more about just filling in the missing pixels in that we've got hardware that is even at the very high end of what's available struggling to really utilize the feature as it is and the feature's implementation is still in my opinion pretty early stages in terms of how useful it is because it's not creating barrels out of nothing that perfectly integrate into the rest of the scene. It's just applying lipstick, you know?
And to be clear, it's not applying that lipstick in a universally unpleasant way. Like I actually found that particularly with background characters where there was less time and attention that could be spent on that, I actually was pretty pleased with the results sometimes. Like I loved the extra little, you know, shot of immersion that I got walking around in a street populated by NPCs in Cyberpunk. And that's something that, you know, I hate to validate the 'well, this was our intent, we just didn't have time to spend on it' argument because I think the flip side of that argument is, well, you know, that's a shortcut, right? Like we're giving game developers the tools to be lazy. I don't buy that. But, you know, the barrel one, I think you could kind of make an argument for, you know, well, yeah, maybe you should draw a barrel. Like, isn't this art? Yeah, right.
But I think, you know, every game ever that I can think of that has a main character has struggled a little bit from the main character looking a little bit out of place next to all the NPCs that are procedurally generated or that are drawn, but obviously they can't spend as long. Or you're in a world where there's like half a dozen different models for your non-playable characters, right? Like there's always a compromise there. And if I could have AI fill in the gaps, you know what? I find that to be an immersion-increasing experience. It's just that we don't have the hardware to run it today and it's more subtle today than I think it's going to need to be to really change my gaming experience enough to justify the performance hit. And it's a big one.
Yeah. Another aspect of the future of DLSS5 will be how they continue to optimize it. Since the spring, they say that it's five times more performant. When they first showed the demos at that time, it required two 5090s to make it work.
My god. We should just be grateful. We should be grateful it doesn't require two 5090s.
Yeah, exactly. Bring it back.
Who knows when we're getting the next generation of graphics cards, but hopefully Nvidia will continue to optimize, make it faster. One thing that I noticed is that they are definitely paying attention to the backlash that's been happening. Not only did they encourage people to post videos earlier to kind of, I think, balance out sort of the fan outcry of DLSS5, they have now said that they are going to work on a 4000-series version of DLSS5. And so we know that they're paying attention. We know that they are trying to mitigate the PR disaster that it's become. And hopefully that means for gamers that want to use it, they won't have to buy 7090s or whatever it will be in three, four years, right?
I mean, I don't have much hope for that. The 40 series can run it, but this is the lightest version of DLSS5 that there will likely ever be. I shouldn't say that this isn't the lightest version, but it feels like we're going to be going down a path where just like the Transformer model and the way that that just doesn't run as well on older series cards, it has a bigger performance hit. I think what we're going to see is we're going to see DLSS5 sort of versions that are tuned for the different generations of cards. And a 40 series just doesn't have the same neural capabilities as a 50 series. And neither is going to be anywhere near what we're going to see in the future. So yeah, you might have the feature, but I just don't think it's realistic that you're going to be running the future variants of this that are going to really meaningfully impact the gaming experience.
But I mean, it's a funny thing that we're even having this conversation because when I first got into this hobby, oh, I don't know, you know, almost 25 years ago now, it was like, well, yeah, you know, your old card doesn't support DirectX 9C, so you don't have the new feature. And everyone was like, okay, I guess I'll need a new one. But then again, back then you could get a new card for 150 bucks that was reasonably performant. So people's mentality around how long they want to keep their cards and have access to all the new features has totally changed. But so has the cost structure of upgrading. It's gotten insane, right?
Like, I'm just going to jump over on Newegg because the last time I was looking at any of this, 50 series was going up in price again, right?
It's because local AI and things like that.
I think... Holy smokes. How much... 5080s are like $1,560 now?
Oh my god. That's how much you can build an entire computer for and less than that.
