Nvidia DLSS Explained: How Upscaling Evolved Into DLSS 5

Nvidia DLSS has been astir since the RTX 2080 dropped backmost successful 2018, but portion it started arsenic a mode to usage instrumentality learning to upscale games, it's grown to beryllium truthful overmuch much than that.

Now, 8 years aft the Tensor Core that powers DLSS archetypal debuted successful Nvidia's Volta GPU architecture, DLSS volition upscale your games, make wholly caller frames, and, erstwhile DLSS 5 comes retired aboriginal this year, volition adjacent re-draw each framework successful your game. Some of these features are much divisive than others, but it's hard to reason that DLSS isn't 1 of the astir important GPU bundle suites successful a while, and is simply a large portion of wherefore Nvidia graphics cards are truthful good.

DLSS Upscaling

DLSS ostensibly stands for “Deep Learning Super Sampling,” and that’s precisely what it did astatine the beginning. The full thought was to render the crippled astatine a little resolution, and past usage an AI algorithm, trained connected Nvidia ace computers and moving connected Tensor Cores, to accurately upscale to a higher resolution.

For astir games, DLSS is disposable successful 1 of 4 presets, each changing the scaling origin of the game.

• Ultra Performance: 33%
• Performance: 50%
• Balanced: 58%
• Quality: 66.7%

Obviously, the higher you spell up, the amended the representation prime is going to be, but that volition conversely impact your framework rate. As a wide regularisation of thumb, I’d urge ‘Performance’ for 4K, ‘Balanced’ for 1440p, and ‘Quality’ for 1080p gaming. I personally crook DLSS connected whenever I can, but that’s due to the fact that the algorithm is truthful bully these days that I can’t archer the quality unless I’m actively looking for it. That wasn’t ever the case, though.

In the beginning, this was rough. The archetypal iteration of DLSS upscaling was noisy and had a ton of issues with artifacts and wide representation softness. Coupled with the information that it lone ran successful a mates of games, and it was hard to instrumentality Nvidia’s upscaling ace seriously. However, with DLSS 2.0, released successful 2020, Nvidia improved the algorithm by taking question vector information from the crippled motor into consideration, making it capable to much accurately make pixels that were existent to the question connected the screen.

DLSS 2 besides shifted Nvidia’s upscaling algorithm from needing to beryllium trained connected each crippled individually, to an algorithm that could beryllium applied to galore antithetic games, assuming developers worked the DLSS files into their game. This made adoption skyrocket, and we started to spot a ton of games starting to enactment the technology.

While the upscaling tech got a small amended implicit the pursuing fewer years, the adjacent biggest leap came with DLSS 4, which debuted with the RTX 5080 successful 2025. This moved the upscaling algorithm from a CNN (convolutional neural network) to a transformer model, which drastically improved accuracy.

This update worked connected each RTX graphics card, and you could adjacent force-enable it successful the operator software, allowing it to tally successful fundamentally immoderate crippled that already supported DLSS 2 oregon 3. The lone downside was that the exemplary was a tad heavier than the older CNN model, truthful it could lessen the show boost you got from enabling DLSS – peculiarly connected older RTX 3000 and 2000 graphics cards.

Then, successful 2026, Nvidia released a weird small mid-generation update, with DLSS 4.5. This changed some upscaling and framework procreation – which I’ll get into aboriginal – but for upscaling it conscionable made the transformer exemplary a spot much close erstwhile you acceptable it to “Performance” oregon “Ultra Performance” modes. The lone occupation was that it accrued the show footprint connected older cards – truthful I’d lone urge enabling it with an RTX 5060 oregon newer graphics card.

Frame Generation

DLSS upscaling has go rather fashionable implicit the past fewer years, but with DLSS 3 and the RTX 4090, Nvidia went from utilizing AI to make other pixels, to generating full frames. Frame gen is incredibly elemental successful theory: It uses an algorithm to make other frames and past inserts them betwixt really rendered frames successful the render queue. However, this is of people a spot much analyzable than it sounds.

By their precise nature, video games are unpredictable, truthful Nvidia had to find a mode to make the other frames that wouldn’t crook games into a messiness of hallucinations and artifacts. And, conscionable similar the upscaling bit, framework procreation was unsmooth astatine first.

Rather than conscionable being escaped show similar Nvidia initially claimed, framework procreation accrued latency, alternatively than lowering it. One of the main purposes of playing games astatine a higher framework complaint is to little the hold betwixt immoderate you’re doing – clicking a mouse, moving your character, oregon immoderate – and that enactment being reflected connected the screen. Because framework procreation isn’t really making the crippled run faster, it tin lone ever adhd latency, due to the fact that the algorithm needs a mates of milliseconds to make each frame.

Nvidia offsets this summation successful latency with Reflex, a strategy that fundamentally zeroes retired the render queue and has the CPU sync with the graphics card, truthful that frames aren’t conscionable sitting determination waiting for the GPU to render them.

Reflex has been retired for years, and it truly does amended latency rather a bit, but it’s indispensable to halt framework procreation from being consecutive up terrible. With framework generation, Nvidia fundamentally added a caller render queue that exists connected the GPU itself, taking a rendered frame, holding it until a caller framework is generated, and past pacing them retired to support the question smooth.

This is besides wherefore framework gen doesn’t precisely treble your framework rate, adjacent though it is generating a caller framework disconnected of each 1 sent by the game. And it’s besides wherefore it hits latency a bit. In my experience, the latency volition typically spell from, for example, 30 sclerosis to 40 sclerosis successful Cyberpunk 2077 erstwhile framework procreation is enabled.

The added latency isn’t ideal, but it’s not capable to beryllium noticeable, particularly successful azygous subordinate games wherever you’re not trying to compete.

As arguable arsenic framework procreation was erstwhile it launched with the RTX 4080, though, Nvidia upped the multiplier with DLSS 4 successful 2025, adding a 3x and 4x framework gen option. Now, RTX 50 bid graphics cards could make up to 3 frames per rendered frame, theoretically boosting framework rates by 4x. Again, that fig doesn’t precisely enactment retired successful theory, but it does vastly amended performance.

Generating each of those rendered frames could person been a nightmare, but Nvidia worked an AMP core, oregon AI Management Processor, into its Blackwell (RTX 50 series) graphics cards. This small spot works arsenic a benignant of taskmaster for the GPU and takes implicit framework pacing, which was typically handled by the CPU previously. Because that enactment is present handled connected the aforesaid dice arsenic the GPU, determination is little latency erstwhile scheduling retired the frames, which is wherefore multi-frame procreation is capable to work.

What’s much impressive, though, is however small the other generated frames interaction latency beyond the archetypal latency interaction of 2x framework gen. In the aforesaid Cyberpunk conception I mentioned earlier, turning framework gen to 4x lone bumps the PC latency up to 43ms, which is barely more. And portion adding latency astatine each is simply a downside, the upside is that it’s overmuch easier to usage a 4K 240Hz show and get your money’s worth.

But 4x framework gen wasn’t capable for Nvidia. At CES 2026, the institution announced DLSS 4.5, a benignant of half-generation refresh of DLSS 4. With it, 6x framework gen is present possible, but the much breathtaking portion of the equation is Dynamic Multi Frame Generation. When enabled, this volition person the GPU, um, dynamically alteration the framework gen multiplier, to support your framework complaint arsenic adjacent to your monitor’s refresh complaint arsenic possible.

Think of it arsenic a benignant of spicy V-Sync, but alternatively of limiting framework complaint to your refresh rate, it conscionable generates other frames to support your show afloat fed. With this update, Nvidia besides updated the AI exemplary that framework procreation runs on. Creatively called “Model B” successful the Nvidia App, it is overmuch amended astatine handling UI elements, which tin messiness up successful immoderate games erstwhile framework gen is active.

Nvidia seems to beryllium heading successful a absorption of perpetually expanding the magnitude of frames that DLSS is going to generate, and who knows wherever it’s going to determine to deed the brakes. But nary substance however precocious that multiplier gets, conscionable support successful caput that this diagnostic decidedly isn’t for anybody.

If you’re inactive rocking a 60Hz display, framework procreation is not going to bash anything for you. And if you’re struggling to deed 50-60 fps successful the archetypal place, framework procreation is going to beryllium a laggy, artifact-riddled mess.

Instead, framework procreation is champion for radical who are already getting 50-60 fps connected a high-refresh display, wherever it’ll marque games look overmuch smoother. Even then, though, the added latency means that a batch of competitory gamers should astir apt conscionable disregard it.

The Future of DLSS

While some DLSS upscaling and framework procreation usage AI to amended performance, they don’t meaningfully alteration the mode the crippled itself looks. While determination mightiness beryllium small mistakes present and there, the last merchandise is highly faithful to what the crippled is outputting. But that mightiness beryllium changing aboriginal this twelvemonth erstwhile DLSS 5 launches.

All we’ve truly seen from DLSS 5 truthful acold is simply a abbreviated demo on-stage astatine GTC 2026 (Graphics Technology Conference), truthful there’s not a batch to spell disconnected of. But, based connected what was there, this caller algorithm seems to massively interaction the aesthetics of a game, peculiarly erstwhile it comes to quality models.

Nvidia claims that the exemplary is grounded successful the geometry and “scene semantics” of the game. But erstwhile the exemplary is taking the game’s last output on with question vector information to make an representation that’s overlaid connected top, it definite looks similar it’s altering the aesthetics of the game.

In theory, this means that the amended the underlying crippled looks, the amended DLSS 5 volition beryllium astatine generating a last image, but due to the fact that it doesn’t instrumentality information from the crippled motor into account, it’s precise imaginable that it volition messiness up. Though, Nvidia volition apt spot it up to marque it look amended implicit the adjacent half-a-year earlier it is disposable to download.

But beyond the interaction to a game’s aesthetic, it remains to beryllium seen whether oregon not it volition really amended performance. After all, since the beginning, 1 of the guiding lights of DLSS was maximising show portion preserving representation quality. The aboriginal exemplary of DLSS 5 needed 2 RTX 5090s to run. And, sure, Team Green volition find a mode to optimize it to tally connected a azygous graphics card, but lone clip volition archer if it’ll beryllium thing that gamers crook on. If it conscionable changes the mode the crippled looks but with worse show – I don’t deliberation a batch of radical are going to take to usage it.

Jackie Thomas is the Hardware and Buying Guides Editor astatine IGN and the PC components queen. You tin travel her @Jackiecobra