DLSS 5 Explained: How Nvidia’s Neural Rendering Is Changing PC Graphics

For most of the history of PC graphics, getting a better looking picture meant one thing: a more powerful graphics card. More shader cores, faster memory, higher clocks. If you wanted to run a new release at higher settings, you upgraded. Then, around 2018, something shifted. Nvidia started using artificial intelligence not just to make games run faster, but to fill in detail the graphics card never actually rendered.

Eight years later that idea has gone much further than most people expected. In March 2026, at its GTC conference, Nvidia announced DLSS 5, which it calls real-time neural rendering. Jensen Huang described it as “the GPT moment for graphics.” Unlike earlier versions of DLSS, which mainly upscaled lower-resolution images and generated extra frames, DLSS 5 uses an AI model to add lighting and material detail to the final image, things like the way light passes through skin, the texture of fabric, or how hair catches light.

It’s also become one of the most argued-about graphics technologies in years. Some reviewers who tried it in NBA 2K27, the first title to ship with it in September 2026, came away impressed. Others found the improvement subtle and the performance and power cost large. Artists and some developers have criticised it for changing the look that studios deliberately designed. Nvidia’s CEO responded by saying the critics were “completely wrong.”

In this guide we’ll explain all of it properly. How a frame of a modern game actually gets drawn, how DLSS evolved from a simple upscaler to neural rendering, what DLSS 5 does differently, what early hands-on testing shows, why people are so divided, what AMD and others are doing, and what it all means if you’re choosing a graphics card or power supply in late 2026.

As always, we’ve linked our sources at the end. Where something is Nvidia’s own claim rather than independent testing, we say so.

The short version, if you’re in a hurry

What it is: DLSS 5 is Nvidia’s newest version of its AI graphics technology. Its headline feature, which Nvidia calls 3D-Guided Neural Rendering, uses an AI model to enhance lighting, skin, hair, fabric and material detail in each frame, while staying tied to the game’s own scene data.
What it isn’t: It’s not mainly about higher frame rates. In early testing, turning on the neural rendering part actually reduced frame rates noticeably. DLSS 5 still includes upscaling and frame generation, which recover performance.
Hardware: At launch it requires a GeForce RTX 50 series card (or an RTX Spark laptop). Nvidia has said support for RTX 40 series is coming later.
Where it works: NBA 2K27 launched with it on September 3, 2026. Hogwarts Legacy gets it on November 5, and Where Winds Meet in late November. Bethesda, Capcom, Ubisoft, Tencent and Warner Bros. Games have said they plan to adopt it.
Early testing: Engadget measured an RTX 5090 at 4K dropping from about 240 to 150 frames per second with DLSS 5 on in NBA 2K27, and power draw rising from around 400 to 525 watts. Visual gains were clearest in close-ups and much less obvious in normal wide shots.
The controversy: Critics say it can override a studio’s art direction, smoothing faces and brightening deliberate shadows. Nvidia says developers keep control and the feature is optional.
For buyers: It’s interesting, but not a reason to rush out and buy a new card today. If you do use it, the extra power draw means your power supply matters more than ever.

What’s in this guide

How a frame gets drawn, and why it’s so expensive
A short history of DLSS, from version 1 to 4.5
What DLSS 5 actually does
Which cards and games support it
What early hands-on testing shows
The art direction controversy
What AMD, Intel and Microsoft are doing
What it means for your next graphics card and power supply
Recommended settings and practical tips
How an AI learns to render
Where neural rendering goes next
What to expect on each RTX 50 card, plus a glossary
Myths and misunderstandings
Frequently asked questions

Part 1: How a frame gets drawn, and why it’s so expensive

To understand why AI has become such a big part of PC graphics, it helps to know what your graphics card is actually doing sixty, a hundred, or two hundred times every second.

Rasterisation: the classic way

A 3D scene in a game is built from millions of tiny triangles, each with textures (images wrapped onto surfaces) and material properties (how shiny, rough or metallic something is). The traditional way to draw a frame is called rasterisation. The graphics card works out which triangles are visible from the camera’s point of view, turns them into pixels, and then runs small programs called shaders on each pixel to work out its colour, taking into account lights, shadows, textures and so on.

Rasterisation is fast, and decades of clever tricks have made it look very good. But many of those tricks are approximations. Reflections, soft shadows, light bouncing off one surface onto another, light scattering through skin, these are all things rasterisation struggles to do accurately, so developers fake them with techniques that look convincing most of the time and slightly off some of the time.

Ray tracing: more accurate, much more expensive

Ray tracing simulates how light actually behaves, by tracing the paths of rays of light through the scene as they bounce off and pass through surfaces. It produces far more realistic lighting, reflections and shadows. Path tracing, the most complete form, traces many bounces for many rays per pixel.

The problem is cost. Fully path tracing a modern game at 4K, at high frame rates, in real time, requires vastly more computing power than any consumer graphics card has. Even the fastest cards can only trace a limited number of rays per pixel per frame, so the raw result is noisy, grainy and incomplete. Something has to clean it up and fill in the gaps.

Resolution makes everything harder

On top of that, displays keep getting sharper and faster. A 4K screen has about 8.3 million pixels, four times as many as 1080p. A 240Hz monitor wants 240 frames every second. Multiply those together and the amount of work grows enormously. Rendering every one of those pixels from scratch at high quality, every frame, simply isn’t practical, even on a top-end card.

Where AI comes in

This is the core idea behind DLSS and everything like it. Instead of rendering every pixel the hard way, the graphics card renders fewer pixels, or fewer frames, or a noisier version of the lighting, and an AI model trained on huge amounts of high-quality images reconstructs the rest. The AI part runs on dedicated hardware inside RTX cards called Tensor Cores, which are very efficient at the kind of math neural networks use.

When it works well, you get an image that looks as good as, or sometimes better than, rendering natively, for much less effort. When it works badly, you get artifacts: smearing, ghosting trails behind moving objects, shimmering, or details that look “invented.” The history of DLSS is basically a story of Nvidia steadily improving the good side and shrinking the bad side, and then, with DLSS 5, trying something much more ambitious.

Part 2: A short history of DLSS, from version 1 to 4.5

DLSS stands for Deep Learning Super Sampling. The name made more sense in the early days, when upscaling was all it did. Today it’s more of an umbrella brand for a whole family of AI graphics features. Here’s how it got here.

DLSS 1 (2018–2019): a rough start

DLSS launched alongside the first RTX cards, the GeForce RTX 20 series, in 2018. The idea was simple: render the game at a lower resolution, then use an AI model to upscale it to your screen’s resolution, so you’d get a sharper image than plain upscaling at a similar frame rate. That would free up performance to turn on ray tracing, which those cards also introduced.

The first version was honestly not great. It needed to be trained separately for each game, image quality was often soft and blurry, and many people preferred to just lower the resolution the old-fashioned way. But it established the concept, and the hardware, Tensor Cores, was now sitting in every RTX card.

DLSS 2 (2020): the version that made people care

DLSS 2 changed the approach. Instead of per-game training, it used a single general model, and it fed the AI more information: not just the current low-resolution frame, but data from previous frames and motion vectors that tell the model how each pixel is moving. This is called temporal upscaling.

The difference was dramatic. In many games, DLSS 2’s “Quality” mode looked about as good as native resolution, sometimes sharper in fine details, while running much faster. This is the point where DLSS went from a curiosity to a feature people actively looked for when choosing a graphics card.

DLSS 3 (2022): frame generation

With the RTX 40 series in 2022, Nvidia added Frame Generation. Instead of only upscaling each rendered frame, the AI generated entirely new frames in between the ones the card actually rendered, using motion data and a dedicated piece of hardware called the optical flow accelerator. In practice this could nearly double the frame rate you see on screen.

Frame Generation came with a catch that’s still important today: latency. The generated frames make motion look smoother, but they don’t make the game respond faster to your mouse or controller, because the game itself is still running at the lower “real” frame rate. In fact, the process adds a little delay. Nvidia pairs it with Reflex, its latency-reduction technology, to compensate. For fast, competitive play, many people still turn frame generation off.

DLSS 3.5 (2023): Ray Reconstruction

Remember how ray tracing produces a noisy, grainy image that needs cleaning up? Traditionally that’s done with hand-written “denoisers.” DLSS 3.5 introduced Ray Reconstruction, which replaced those denoisers with an AI model trained to turn noisy ray-traced lighting into clean, detailed lighting. It made ray-traced reflections and lighting noticeably sharper and more stable in supported games, and it worked on all RTX cards.

DLSS 4 (2025): Multi Frame Generation and the transformer model

Launched with the RTX 50 series in early 2025, DLSS 4 brought two big changes. First, Multi Frame Generation, which generates several AI frames for every rendered frame instead of just one, on RTX 50 cards. That’s how Nvidia arrived at some of its very high frame rate claims for the RTX 50 series. Second, it switched the AI models behind upscaling and Ray Reconstruction to a new architecture called a transformer, the same broad type of neural network behind modern AI chatbots. The transformer models produced noticeably better image stability and detail, and they worked on older RTX cards too.

DLSS 4.5 (January 2026): the second-generation transformer

At CES in January 2026, Nvidia announced DLSS 4.5. It introduced a second-generation transformer model for Super Resolution and DLAA (Nvidia’s anti-aliasing mode that runs the AI at native resolution), which Nvidia says uses five times the compute of the original transformer model. It also introduced Dynamic Multi Frame Generation, which adjusts how many frames are generated to match your display’s refresh rate, with modes reported to go up to 6x on RTX 50 series cards. Nvidia said the upgraded Super Resolution model was available across hundreds of games and apps, and worked on RTX 20, 30, 40 and 50 cards.

DLSS 4.5 arriving in January makes the timing of DLSS 5 all the more striking. Just two months later, Nvidia was announcing an entirely new direction.

A quick summary

Version	Year	Main idea
DLSS 1	2018	AI upscaling, trained per game
DLSS 2	2020	General temporal upscaling using motion vectors
DLSS 3	2022	Frame Generation (RTX 40 and newer)
DLSS 3.5	2023	Ray Reconstruction, AI denoising for ray tracing
DLSS 4	2025	Multi Frame Generation (RTX 50), transformer models
DLSS 4.5	2026	Second-gen transformer, Dynamic Multi Frame Generation
DLSS 5	2026	3D-Guided Neural Rendering of lighting and materials

Part 3: What DLSS 5 actually does

Every version of DLSS up to 4.5 was, at its heart, about reconstruction: taking incomplete information (fewer pixels, fewer frames, noisy rays) and reconstructing what the full, correct image should look like. The goal was to match what the game would look like if you had infinite computing power.

DLSS 5’s headline feature goes a step further. It doesn’t just reconstruct the image the game was trying to draw. It adds visual detail that the game’s renderer didn’t compute, particularly in how light interacts with materials.

The neural rendering pass

According to coverage of Nvidia’s announcement, the DLSS 5 model takes the colour and motion vector data from each frame and applies a neural rendering pass on top. That pass is designed to improve things that are notoriously hard and expensive to render properly in real time:

Skin. Real skin is slightly translucent. Light enters it, scatters around under the surface, and comes back out, giving that soft, warm glow faces have in good lighting. Games usually approximate this, often leaving characters looking a bit like plastic or wax.
Hair. Thousands of thin, semi-transparent strands catching light from different angles is very difficult to render convincingly.
Fabric. Cloth has fine weave patterns and a soft sheen that changes as it moves.
Light-material interaction in general. How light falls across different surfaces, subtle shadowing in crevices, the way rough and smooth materials catch highlights.

Nvidia calls it “3D-Guided” because the AI isn’t just painting over a flat image. It’s guided by the game engine’s scene data, so its output stays tied to the actual geometry and materials in the scene. Nvidia says the result is deterministic and stable from frame to frame, meaning the same scene produces the same output and details don’t flicker or “boil” as the camera moves, which has been a common problem with generative AI imagery.

Developer controls

Nvidia says developers get controls over how strong the effect is, including intensity, colour grading and masking options. Coverage has described two sliders in particular: “Structure Intensity,” which controls high-frequency details such as ambient occlusion (the soft shadowing in corners and crevices) and reflections, and “Tone Intensity,” which adjusts broader colour and lighting responses. Masking means developers can exclude parts of the image, for example UI elements or specific characters, from being processed.

These controls are central to Nvidia’s response to criticism, as we’ll see later. In theory, a studio can dial the effect up for photorealistic scenes and down, or off, where it clashes with its intended look.

Still includes the rest of DLSS

It’s worth being clear: DLSS 5 is a suite, not a single feature. Nvidia lists Super Resolution, Ray Reconstruction, Dynamic Multi Frame Generation and 3D-Guided Neural Rendering as DLSS 5 technologies. The upscaling and frame generation parts still exist and still improve performance. It’s specifically the neural rendering part that costs performance rather than saving it, which is why you’ll often see it used together with frame generation.

Part 4: Which cards and games support it

Graphics cards

At launch, the neural rendering part of DLSS 5 requires a GeForce RTX 50 series graphics card. That includes desktop cards from the RTX 5050 and 5060 family up to the RTX 5090, and RTX 50 laptop GPUs. Nvidia has said DLSS 5 runs up to 4K resolution.

Nvidia’s RTX Spark laptop chip, which we covered in our previous article, also supports DLSS 5, since it uses the same Blackwell graphics architecture as the RTX 50 series.

For RTX 40 series owners there’s some good news: Engadget reported that Nvidia has announced DLSS 5 will eventually support RTX 40 cards. No date was given at the time of writing. Older RTX 20 and 30 cards aren’t included in the neural rendering part, although they still benefit from the improved DLSS 4.5 Super Resolution model.

Why only newer cards? The neural rendering model is heavy. It runs a substantial AI network on every frame, and it relies on the Tensor Core capabilities and FP4 support in newer architectures. Even on an RTX 5090, as we’ll see, it has a big performance cost.

Games

Here’s the current picture:

NBA 2K27 launched with DLSS 5 support on September 3, 2026, the first official title to ship with it.
Hogwarts Legacy is getting a DLSS 5 update on November 5, 2026, according to Nvidia.
Where Winds Meet is adding DLSS 5 in late November, per Nvidia.
Publishers committed: Bethesda, Capcom, Ubisoft, Tencent and Warner Bros. Games were named as planning to adopt the technology, though specific titles haven’t all been announced.

Engadget also noted that people with recent Nvidia cards have been modding DLSS 5 into existing titles like Control and Cyberpunk 2077. That’s interesting as an experiment, but unofficial mods don’t have the developer tuning that the official implementations get, and the results can look very different. That may partly explain some of the stranger looking clips that circulated online.

Nvidia’s performance claims

Nvidia’s own marketing focuses on the whole DLSS stack. Wccftech, reporting from a hands-on event, passed along Nvidia’s claim that NBA 2K27 ran at 227 frames per second with DLSS off versus 370 with DLSS on, using the full DLSS stack including frame generation. Nvidia also says the latest updates to NBA 2K27, Hogwarts Legacy and Where Winds Meet run its models up to “2X faster.” These are manufacturer figures under Nvidia’s chosen conditions, and they include frame generation, so they don’t tell you what neural rendering on its own costs.

Part 5: What early hands-on testing shows

This is where things get more interesting than the marketing. Several outlets tested DLSS 5 in NBA 2K27, both at preview events and after the game launched. Their findings line up in some areas and differ in others.

The visuals: real, but mostly in close-ups

Most testers agree the neural rendering is doing something real. Engadget, testing on an RTX 5090 at 4K, found it added detail to player and crowd models, with skin texture improving especially on faces and arms, and shadows looking slightly more realistic. Without DLSS 5, the author described the models as looking like “plastic mannequins.” Other hands-on reports described sharper facial structure, cleaner hair and more convincing skin detail.

But there’s an important caveat that came up repeatedly: the effect is strongest in close-up shots. Engadget saw no major benefit in the wide overhead view you spend most of a basketball match looking at, and described the effect as “almost too subtle” for most players, less noticeable than ray tracing. GamesRadar’s hands-on found the effect didn’t over-beautify characters in normal play, but noticed a lip-sync mismatch in dialogue scenes.

TechRadar’s reporter came away from a hands-on at Gamescom genuinely impressed, saying the experience turned them around on the technology, and that Nvidia explained why the initial reveal images looked so overdone compared with the shipping implementation.

The performance cost: significant

Here’s where opinions shift. The neural rendering pass is expensive.

Engadget: on an RTX 5090 at 4K, frame rate dropped from about 240 frames per second to about 150 with DLSS 5 on.
Other reviewers of the final release reported drops of roughly half or more at 4K on the highest settings, with frame generation used to recover smoothness at the cost of extra latency.
TechRadar reported that a laptop with an RTX 5060 ran smoothly with DLSS 5 enabled, using 2x frame generation, which suggests it’s usable beyond the very top-end cards, at least in this particular title.

The differences between testers come down to different settings, scenes, resolutions and whether frame generation was on. The common thread is clear though: neural rendering trades performance for image quality. It’s the opposite of what most people expect from the DLSS name.

The power cost: also significant

Engadget measured something that deserves more attention than it got. With DLSS 5 off, their RTX 5090 drew about 400 watts. With DLSS 5 on, it rose to about 525 watts. The author also linked a forum user’s report of over 600 watts and a melted power cable, which Engadget couldn’t verify.

We want to be careful not to spread scare stories based on one unverified forum post. But the measured jump from 400 to 525 watts is real, and it makes sense: neural rendering keeps the Tensor Cores busy on top of everything else the card is doing. If you have a high-end card, that has practical implications for your power supply and cabling, which we’ll come back to in Part 8.

Engadget’s verdict, and a fair counterpoint

Engadget’s headline summed up its view: DLSS 5 adds subtle details to NBA 2K27, but demands a lot more power. The subtitle was blunter: “The juice isn’t worth the squeeze.” The author called it a work in progress and far from essential, and even suspected the game’s base character models might have been left a bit undercooked to make the comparison look better.

Thats one fair perspective. Another is that this is the very first shipping implementation, in one title, of a technology that’s months old. DLSS 1 was widely disliked too, and DLSS 2 turned it into something people now consider essential. Whether DLSS 5 follows the same path depends on how quickly Nvidia can reduce its cost and how carefully developers tune it.

Part 6: The art direction controversy

No graphics technology in recent memory has caused as much heated debate as DLSS 5. To understand why, you have to look past the performance numbers to a more basic question: who decides what a game should look like?

What critics are saying

When Nvidia first showed DLSS 5 in March 2026, the reaction from many artists and developers was harsh. The core complaint was that the AI model doesn’t just enhance an image, it changes it, and in ways the people who made the game didn’t choose.

Several specific criticisms came up again and again:

Overriding deliberate lighting. Some reports said the model brightened shadows and flattened contrast that artists had placed on purpose. In games, a dark corner or a harsh shadow is often a creative choice, to create tension, guide your eye, or set a mood.
“Beautifying” faces. Critics said early demos smoothed and idealised characters’ faces, with clips from Resident Evil Requiem and Hogwarts Legacy drawing the most criticism. A character designed to look tired, rough or unsettling might come out looking like a glossy advert.
Homogenisation. A broader worry is that if an AI model trained on photorealistic imagery is applied across many games, different art styles get pulled towards the same statistical idea of “realistic,” and games start to look alike.
Respect for the craft. Concept artist Karla Ortiz was among those who called the approach disrespectful to developers’ intentional art direction. Some critics went as far as labelling the results “AI slop,” a term that’s become shorthand for low-effort, generic AI-generated content.

Nvidia’s response

Nvidia pushed back hard. At GTC, Jensen Huang said the critics were “completely wrong.” His argument was that developers keep direct control over geometry, textures and the generative elements, and he described the approach as “content-control generative AI,” as opposed to generative AI that invents freely.

Nvidia and partners such as Bethesda also stressed that the early demos were just that, early, that developers have artistic control through the intensity, colour and masking settings, and that the feature is optional for players. Reports on Nvidia’s own technical paper said the current version is focused on photorealistic rendering, which suggests it’s not intended for heavily stylised art at all, at least for now.

Not everyone found Huang’s tone persuasive. A TechRadar commentary piece argued the “you’re wrong” response was poor strategy, and questioned what “content-control” really means in practice.

Our take

We think both side have a point, and the truth will depend heavily on how each studio uses it.

The critics are right that applying an AI “realism” pass on top of an artist’s work, without care, can damage what made that work special. Not every game should look photorealistic. Stylised games, from cel-shaded adventures to moody horror, rely on deliberate choices that a generic model could easily undo. And the reveal images clearly overdid it, as even Nvidia’s representatives seem to have acknowledged in later hands-on sessions.

On the other hand, if developers really do control the intensity and masking, then DLSS 5 is a tool, not a filter imposed from outside. In the right hands, used subtly on skin, hair and fabric in a realistic game, the hands-on reports suggest it can genuinely make characters look less like plastic. The fact that the shipping implementation in NBA 2K27 was described by several testers as subtle rather than garish is arguably a sign of a studio using it carefully.

The key thing for players is that it’s optional. If you don’t like what it does in a particular game, turn it off. If it becomes something developers can’t opt out of, or that changes art without their consent, that would be a very different conversation.

Part 7: What AMD, Intel and Microsoft are doing

Nvidia isn’t the only company betting on AI in graphics. The whole industry is moving in this direction, although nobody else has announced something quite like DLSS 5’s neural rendering pass yet.

AMD: FSR Redstone and FSR 4

AMD’s equivalent technology is called FSR, FidelityFX Super Resolution. For years, FSR used hand-written algorithms rather than AI, which made it work on almost any graphics card but generally left it behind DLSS in image quality. That changed with FSR 4, which uses machine learning for upscaling on AMD’s Radeon RX 9000 series cards.

AMD has since brought its ML-based features together under an umbrella called FSR Redstone. According to coverage of the launch, Redstone includes AMD’s ML upscaling (with FSR 4 now supported in over 200 games), ML frame generation, “ray regeneration” (AMD’s answer to Ray Reconstruction for cleaning up ray-traced lighting), and radiance caching, a technique for more efficient lighting that’s slated to reach developers in 2026. The ML features require Radeon RX 9000 series cards. AMD has also said it plans to bring a lightweight version of FSR 4 to handheld PCs before 2027.

There have been reports that AMD intends to open FSR 4 upscaling up to run on competitors’ hardware, including Nvidia cards, though accounts of the exact requirements differ.

Intel: XeSS

Intel’s upscaler, XeSS, uses AI and runs best on Intel’s own Arc graphics, with a fallback mode for other cards. It’s a solid technology, but adoption in games has lagged behind DLSS and FSR. Intel hasn’t announced anything comparable to neural rendering.

Microsoft: making AI graphics a standard

One of the long-term problems with DLSS is that it’s proprietary. Developers have to implement Nvidia’s version, AMD’s version and Intel’s version separately. Microsoft has been working on standardising pieces of AI-assisted rendering in DirectX, its graphics API. A feature called Cooperative Vectors, discussed by AMD in the context of Redstone, is designed to let neural networks run efficiently on the dedicated AI units in modern GPUs through DirectX, so developers can write neural shaders once and run them on hardware from different vendors.

If that standardisation takes hold, it could matter more in the long run than any single company’s branded feature. It would mean neural rendering techniques become a normal part of how games are made, rather than a vendor-specific extra.

Why everyone is heading this way

The underlying reason is physics and economics. Making graphics chips dramatically faster through raw transistor count is getting harder and more expensive with each generation, and as we covered in our article on the 2026 memory crisis, the memory that graphics cards depend on is in short supply. AI techniques let companies squeeze much more visual quality out of the same hardware. Whatever you think of DLSS 5 specifically, AI-assisted rendering is clearly the direction the whole industry is heading.

Part 8: What it means for your next graphics card and power supply

So, practically speaking, should DLSS 5 change what you buy? Heres our honest view.

Don’t buy a card just for DLSS 5, yet

Right now, DLSS 5’s neural rendering is available officially in one title, with two more arriving in November. The effect is mostly visible in close-ups, it costs a lot of performance, and the technology is clearly still maturing. That’s not a good reason on its own to spend a large amount of money on a new graphics card, especially in a year when, as we covered in our memory article, graphics card prices have been pushed up by memory shortages.

If you’re already buying a new card for other reasons, though, DLSS 5 support is a genuine bonus for an RTX 50 series card, and the improved DLSS 4.5 upscaling and frame generation are useful features in their own right.

If you have an RTX 40 card, wait

Nvidia has said RTX 40 support for DLSS 5 is coming. If you own an RTX 4070, 4080 or 4090, there’s very little reason to upgrade just for DLSS 5. Wait and see how it performs on your card when support arrives.

VRAM still matters

AI models need memory, and DLSS features use some VRAM themselves. Combined with ever-larger textures in modern titles, this is one more reason not to buy a graphics card with too little VRAM in 2026. For 1440p, 12GB is a sensible minimum, and 16GB gives you more breathing room for high texture settings plus AI features. For 4K, 16GB or more is the comfortable target. We’ll go into this in much more detail in our upcoming guide to buying a graphics card in late 2026.

Your power supply matters more than ever

This is the practical point we think gets overlooked. Engadget measured an RTX 5090’s power draw rising from about 400 to about 525 watts with DLSS 5 on. Neural rendering keeps more of the GPU busy, so the card draws closer to its maximum rated power more of the time. That has a few implications:

Size your power supply properly. Nvidia’s official board power ratings are roughly 575W for the RTX 5090, 360W for the RTX 5080, 300W for the RTX 5070 Ti and 250W for the RTX 5070. Nvidia recommends a 1000W power supply for the RTX 5090 and 850W for the RTX 5080. Don’t plan your PSU around a card’s “typical” power draw. Plan it around the maximum, plus your CPU, plus headroom.
Buy quality, not just wattage. A good ATX 3.0 or 3.1 power supply from a reputable brand handles the short power spikes modern cards produce far better than a cheap unit with an inflated label. When a graphics card costs as much as it does in 2026, saving money on the PSU is a false economy.
Take care with the 12V-2×6 connector. High-end RTX 50 cards use the 16-pin 12V-2×6 power connector (an updated version of 12VHPWR). Make sure it’s pushed in fully and evenly until it clicks, avoid sharp bends right next to the connector, and use the cable that came with your power supply or a quality native cable rather than cheap adapters. Most problems reported with these connectors over the years have been linked to poor seating or bending.
Think about cooling. More power means more heat. Make sure your case has decent airflow, especially with a top-end card.

If you’re not sure whether your current power supply is up to the job, its worth checking before you enable heavy features like DLSS 5 on a high-end card. We’re always happy to look at a PSU model and tell you honestly if it’s adequate.

For laptop users

RTX 50 series laptops support DLSS 5 too, and TechRadar’s hands-on suggested even an RTX 5060 laptop could handle it in NBA 2K27 with frame generation. Just remember that laptops have fixed power and thermal limits, so the performance cost of neural rendering may be felt more heavily, and fans will likely work harder.

Part 9: Recommended settings and practical tips

If you have a supported card and game, here’s how we’d approach the settings.

Start with upscaling

DLSS Super Resolution (upscaling) is the part that almost everyone should use where it’s available. The current transformer models look very good in “Quality” mode, and often fine in “Balanced.” Use “Performance” mode at 4K, where the higher output resolution gives the AI more to work with. DLAA (running the AI at native resolution for anti-aliasing) is great if you have performance to spare.

Use frame generation thoughtfully

Frame generation makes motion look smoother but adds a little latency, so it works best when your base frame rate is already reasonable, roughly 50 to 60 frames per second or more before generation. If your base frame rate is very low, frame generation can make the game look smooth but feel sluggish. Always pair it with Reflex. For fast, competitive multiplayer, many players prefer to leave frame generation off entirely and prioritise responsiveness.

Try neural rendering in slower, story-driven titles

DLSS 5’s neural rendering makes most sense where you’ll actually notice it: slower-paced, cinematic, character-focused games with lots of close-ups. In a fast competitive match, you’re unlikely to notice subtle skin detail, but you will notice the frame rate drop. Try it on and off in the same scene and decide for yourself whether the difference is worth the cost. There’s no wrong answer.

Watch for artifacts

With any AI rendering feature, watch for smearing behind moving objects, shimmering on fine details, or anything that looks “off” about faces. If you see problems, try a different DLSS mode or update the game and driver, since Nvidia and developers regularly ship improved models.

Keep your drivers updated

New DLSS models and fixes usually arrive through driver updates and the Nvidia app, which can also override the DLSS model version used by a game. Keeping your driver current is the easiest way to get better image quality over time.

Monitor temperatures and power

If you’re on a high-end card and turn on DLSS 5, keep an eye on GPU temperature and power draw for the first few sessions using a monitoring tool. If your system becomes unstable or shuts down under load, that’s often a sign the power supply is struggling.

Part 10: How an AI learns to render

A lot of the confusion around DLSS 5 comes from not knowing how these AI models are made in the first place. You don’t need a computer science degree to get the basic idea, and understanding it makes both the excitement and the criticism easier to follow.

Training vs running

Every AI model has two very different stages. First there’s training, where the model learns. Then there’s inference, where the trained model is actually used to do its job. Training happens once (or occasionally, when the model is updated), on huge computers. Inference happens on your graphics card, every single frame, millions of times over the life of a game.

For DLSS, Nvidia does the training on its own supercomputers in its data centres. The finished model is then shipped to your PC through driver updates or with the game. When you turn on DLSS, your card’s Tensor Cores run that model on each frame. Your PC isn’t learning anything while you play, it’s applying what was learned elsewhere.

What the model learns from

For upscaling, the training idea is fairly intuitive. Nvidia renders scenes at very high quality, sometimes at extremely high resolutions with heavy supersampling, which takes far longer than real time. Those become the “ground truth,” the ideal answer. It also renders the same scenes at low resolution, the way your card would in real time. The model is shown the low-quality version, along with motion vectors and other data, and asked to produce the high-quality version. Its output is compared with the ground truth, and the model is adjusted slightly to reduce the difference. Repeat that millions of times across countless scenes, and the model gets very good at predicting what a high-quality frame should look like from limited information.

For neural rendering, the principle is similar, but the “ideal answer” is different. Instead of just a sharper version of the same image, the target includes more realistic lighting and material behaviour, for example how skin should look with proper subsurface scattering, computed with much more expensive offline rendering techniques. The model learns to predict that richer result from what the game’s real-time renderer produces.

Why this matters for the art debate

This is exactly where the art direction controversy comes from. A model learns whatever its training targets teach it. If the targets are photorealistic, the model learns to push images towards photorealism. Nvidia’s own technical description reportedly says the current version focuses on photorealistic rendering. That’s fine for a game aiming at realism, but for a game with a deliberately different look, a model with a strong built-in idea of “correct” lighting could fight against the artist’s choices.

That’s also why Nvidia emphasises developer controls. Intensity sliders and masking are, in effect, ways for developers to limit how much of the model’s learned “taste” gets applied, and where.

Why it’s stable, or not

Generative AI images you might have seen online, the kind created from a text prompt, can produce something slightly different every time, and in video they often flicker or morph between frames. That would be unacceptable in a game. Nvidia says DLSS 5’s output is deterministic and temporally stable, meaning the same input gives the same output, and details stay consistent as you move. It achieves that partly by anchoring the model to the game’s 3D data and motion vectors, rather than letting it generate freely. Early hands-on reports didn’t highlight flickering as a major problem, though the lip-sync issue GamesRadar noticed shows there are still rough edges.

Part 11: Where neural rendering goes next

DLSS 5 is unlikely to be the end of this road. Here are a few directions the industry is clearly exploring, based on what Nvidia, AMD and Microsoft have talked about. Some of these are already partly here, some are further away.

Neural shaders and materials

Rather than applying one big AI pass to the whole frame at the end, future games may use small neural networks inside the rendering process itself, for individual materials or effects. A complex material like layered car paint, velvet, or wet stone could be represented by a tiny trained network instead of a long hand-written shader. Microsoft’s work on Cooperative Vectors in DirectX is designed to make exactly this kind of thing run efficiently across different vendors’ hardware.

Neural texture compression

Textures take up a huge amount of VRAM, and as we’ve covered, VRAM is expensive and limited. AI-based texture compression aims to store textures in a much more compact form and reconstruct them on the fly. If it works well and becomes widespread, it could ease VRAM pressure for a lot of people, which would be very welcome in a world of expensive memory.

Better path tracing at lower cost

Ray Reconstruction already uses AI to clean up noisy ray-traced lighting. AMD’s radiance caching in FSR Redstone is another approach to making realistic lighting cheaper. The long-term goal across the industry is to make fully path-traced lighting practical on mainstream cards, with AI doing much of the heavy lifting.

Lower cost for DLSS 5 itself

Nvidia has a strong track record of making each DLSS feature cheaper and better after launch. Its claim that updated versions of NBA 2K27, Hogwarts Legacy and Where Winds Meet run the models up to twice as fast suggests that process has already started. Support for RTX 40 cards, which Nvidia has announced, will also require it to run acceptably on less powerful hardware.

The open question: who controls the look?

As these techniques spread, the question raised by DLSS 5’s critics will only get more important. If AI becomes a core part of how every frame is produced, then the training data and the controls given to artists become creative decisions in their own right. The healthiest outcome is one where AI rendering tools are just that, tools, used deliberately by artists, rather than a layer of “realism” applied to everything by default.

Part 12: What to expect on each RTX 50 card, plus a glossary

“Supports DLSS 5” means very different things on a budget card and a flagship. Here’s a rough, practical guide to how we’d think about DLSS features across the RTX 50 desktop range. This is general guidance based on each card’s class and the early testing we’ve described, not precise benchmarks, which will vary a lot by game and settings.

RTX 5050 and RTX 5060 (entry level)

These cards are aimed at 1080p. For them, DLSS Super Resolution is the most valuable feature by far, turning a borderline frame rate into a comfortable one at 1080p, or making 1440p possible in lighter titles. Frame generation can help smoothness in slower-paced games, as long as the base frame rate isn’t too low. Neural rendering is technically supported, and TechRadar’s hands-on with an RTX 5060 laptop was positive in NBA 2K27, but in heavier games we’d expect most owners to prioritise frame rate over the extra detail. Also watch VRAM: 8GB versions of entry cards can run short in demanding new releases at high texture settings.

RTX 5060 Ti and RTX 5070 (mainstream)

The sweet spot for 1440p. Upscaling in Quality or Balanced mode gives very good results here, and frame generation works well when your base frame rate is in a healthy range. Neural rendering becomes a realistic option in slower, cinematic games, especially paired with frame generation. Choose the 16GB version of the 5060 Ti if you can, extra VRAM gives you more headroom for AI features and high-resolution textures in the years ahead.

RTX 5070 Ti and RTX 5080 (high end)

Strong 1440p and capable 4K cards with 16GB of VRAM. These have enough headroom to try neural rendering in supported titles while keeping a good experience, particularly with upscaling and frame generation. Power draw is significant, so make sure your PSU is a quality unit in the 750 to 850W range or more, depending on your CPU and Nvidia’s recommendation for the specific card.

RTX 5090 (flagship)

The card most early DLSS 5 testing was done on. It handles neural rendering at 4K, but even here the frame rate drop is large, from about 240 to about 150 frames per second in Engadget’s test, and power draw rises substantially. If you own one, a high-quality 1000W-class power supply, careful 12V-2×6 cable seating and good case airflow are essential, not optional.

If you’re on RTX 40 series

Use DLSS 4.5 Super Resolution and Frame Generation, which already work well on your card, and wait for Nvidia’s promised DLSS 5 support to see how neural rendering performs. There’s no need to upgrade for it.

If you’re on RTX 20 or 30 series

You still benefit from the improved transformer upscaling models, which can noticeably improve image quality compared with older DLSS versions. Frame generation and neural rendering aren’t available. If your card is still giving you the experience you want, DLSS 5 alone isn’t a reason to replace it.

Troubleshooting common DLSS problems

Even with a supported card and game, DLSS features don’t always behave the way you expect. Most of the problems we hear about from customers fall into a handful of categories, and almost all of them can be fixed without spending any money.

The DLSS option is greyed out or missing

This is the most common question by a long way. Start with the basics. Check that your graphics driver is current, because new DLSS features often need a recent driver before the game will even show the option. Then check that the game itself is updated, since DLSS 5 support in titles like Hogwarts Legacy arrives as a patch, not in the original install. Some games also hide frame generation unless Hardware-accelerated GPU Scheduling is turned on in Windows graphics settings, so it’s worth checking that toggle too. If you’re on a laptop, make sure the game is actually running on the Nvidia GPU and not on the integrated graphics, which happens more often than people think.

The image looks soft or blurry

Usually this is a mode issue. Performance and Ultra Performance modes render at a much lower internal resolution, and while the transformer models do a remarkable job, there’s only so much detail the AI can recover at 1080p output from a very low starting point. Try Quality mode, or DLAA if you have the headroom. Also check that the game’s own sharpening slider isn’t set to zero if the developer expects some sharpening, or turned all the way up, which can create halos around edges. A bit of trial and error here goes a long way.

Ghosting or smearing behind moving objects

This tends to come from motion vectors that the game isn’t providing accurately for certain objects, like particle effects, UI elements or fast moving small objects. Newer DLSS models have reduced it a lot. The Nvidia app lets you override which DLSS model a game uses, and switching to the latest preset sometimes fixes ghosting in older titles straight away. If it doesn’t, it’s often a game side bug that only a patch can fix.

The game feels laggy with frame generation on

Remember that frame generation improves how smooth the motion looks, not how quickly the game responds to your input. If your base frame rate before generation is around 30 frames per second, the generated output might read as 100 or more on the counter, but the game will still feel like it’s responding at 30. The fix is to lower other settings or use a stronger upscaling mode first, so the base frame rate climbs, and only then add frame generation on top. Also confirm Reflex is on, since it’s designed to claw back some of that latency.

Crashes or black screens after enabling DLSS 5

Because neural rendering pushes the GPU harder and draws more power, it can expose problems that were hiding before. An unstable overclock, especially a memory overclock, that seemed fine in lighter loads might fall over once DLSS 5 is active. Return the card to stock settings first and test again. If the whole PC shuts off instantly under load, that points strongly at the power supply tripping its protection, which takes us back to the advice in Part 8. And if you’ve had the side panel off recently, double check the 12V-2×6 connector is fully seated.

A note for buyers in Pakistan

We’re a Pakistan based shop, so it’s worth adding some local context that international reviews don’t cover.

Power conditions matter. Voltage fluctuations and load shedding are a fact of life in many cities. A high-end card running DLSS 5 at 500 watts or more is a big, sudden load, and a cheap power supply struggling with unstable mains is not a combination you want. A quality PSU with proper protections, paired with a decent UPS or stabiliser sized for your system’s real peak draw, protects a very expensive investment. If you’re running on a UPS, check its continuous rating can actually handle your full system under load, not just at idle. Many home UPS units that are fine for a basic office PC will cut out with a modern high-end graphics card.

Heat matters too. Summer room temperatures here are far higher than the 20 to 25 degree conditions most international reviews are tested in. A card that runs comfortably in an air-conditioned review lab will run noticeably hotter in a warm room, and DLSS 5’s extra power draw adds to that. Good case airflow, clean dust filters and sensible fan curves make a real difference to both noise and the lifespan of the components.

Price versus benefit. With graphics card prices in Pakistan pushed up by the memory shortage and exchange rates, the gap between tiers is often larger in rupees than the international dollar prices suggest. For most people, a well-chosen mainstream card using DLSS Super Resolution and frame generation will give a much better experience per rupee than stretching for a flagship just to try neural rendering. If you’re unsure where your money is best spent, ask us. We’d rather recommend a sensible build than sell you a card that doesn’t suit how you actually use your PC.

Warranty and genuine stock. Whatever card you buy, make sure it comes with a proper local warranty and is genuine stock. High-end cards are expensive enough that a warranty claim can be the difference between a minor annoyance and a very costly mistake.

Our quick decision guide

You have an RTX 50 card: update your driver, try DLSS 5 in supported titles, keep it for slower cinematic games, and keep an eye on power and temperatures.
You have an RTX 40 card: enjoy DLSS 4.5 and wait for official DLSS 5 support before deciding anything.
You have an RTX 20 or 30 card: use the newer upscaling models through the Nvidia app, and upgrade only when performance in the software you use actually falls short.
You’re buying new: prioritise VRAM and a quality PSU, treat DLSS 5 as a bonus, not the main reason.
You’re on AMD: FSR 4 and Redstone on RX 9000 cards are genuinely good, and you’re not missing anything essential yet.

A quick glossary

DLSS (Deep Learning Super Sampling): Nvidia’s family of AI graphics technologies, including upscaling, frame generation, ray reconstruction and now neural rendering.
Super Resolution: rendering at a lower resolution and using AI to upscale to your screen’s resolution.
DLAA (Deep Learning Anti-Aliasing): running the AI model at native resolution purely to improve edge quality and stability.
Frame Generation: AI creating extra frames between rendered ones for smoother motion. Multi Frame Generation creates several per rendered frame.
Dynamic Multi Frame Generation: adjusting how many frames are generated to match your display’s refresh rate.
Ray Reconstruction: AI cleaning up noisy ray-traced lighting, replacing traditional denoisers.
Neural rendering: using AI to compute or enhance parts of the image itself, such as lighting and material detail, rather than just reconstructing missing pixels or frames.
Tensor Cores: dedicated hardware in RTX GPUs designed to run AI calculations efficiently.
Transformer model: a type of neural network architecture, also used in AI chatbots, that DLSS adopted from DLSS 4 onwards for better image quality.
Motion vectors: data from the game engine describing how each pixel is moving between frames.
Latency: the delay between your input (mouse, keyboard, controller) and the result on screen.
Reflex: Nvidia’s technology for reducing system latency, often paired with frame generation.
VRAM: the dedicated memory on a graphics card.
12V-2×6: the 16-pin power connector used by high-end RTX 50 cards, an updated version of 12VHPWR.
FSR: AMD’s equivalent family of upscaling and frame generation technologies. FSR 4 uses machine learning on Radeon RX 9000 cards.
XeSS: Intel’s AI upscaling technology.
Cooperative Vectors: a DirectX feature designed to let neural networks run efficiently on GPU AI hardware across vendors.

Part 13: Myths and misunderstandings

“DLSS 5 makes games run faster.” The DLSS suite as a whole still includes upscaling and frame generation, which improve performance. But the new neural rendering part reduces frame rates, sometimes by a lot. If you turn on neural rendering expecting free performance, you’ll be disappointed.

“It’s just a filter like the ones on phone cameras.” Not exactly. It’s guided by the game engine’s 3D scene data, including geometry and motion vectors, and is designed to be stable from frame to frame. That’s quite different from a 2D filter painted over the final image. Whether the result always respects the art direction is a separate question.

“It works on all RTX cards.” At launch, the neural rendering part requires an RTX 50 series card or RTX Spark. RTX 40 support has been announced for later. RTX 20 and 30 cards aren’t included, though they still get improved upscaling models.

“It’s forced on you.” No. It’s an option in supported games’ settings. You can turn it off.

“Developers have no say.” According to Nvidia, developers control the intensity, colour grading and masking. How much they use those controls varies by studio.

“Frame generation makes games more responsive.” Frame generation makes motion smoother, but it doesn’t reduce input lag. It adds a little. Responsiveness depends on the real, rendered frame rate.

“You need a 1000W power supply for any RTX 50 card.” No. Power needs vary hugely across the range. A 1000W unit is Nvidia’s recommendation for the RTX 5090. Mid-range cards need much less. Size your PSU for your specific card and CPU, with sensible headroom.

Part 14: Frequently asked questions

What is DLSS 5?

DLSS 5 is the latest version of Nvidia’s AI graphics technology, announced at GTC on March 16, 2026. Its main new feature, 3D-Guided Neural Rendering, uses an AI model to enhance lighting and material detail such as skin, hair and fabric in each frame. It also includes Super Resolution upscaling, Ray Reconstruction and Dynamic Multi Frame Generation.

Which graphics cards support DLSS 5?

GeForce RTX 50 series desktop and laptop GPUs, and Nvidia’s RTX Spark laptop platform. Nvidia has said RTX 40 series support is coming later.

Which games support DLSS 5?

NBA 2K27 launched with it on September 3, 2026. Hogwarts Legacy gets it on November 5, and Where Winds Meet in late November. Bethesda, Capcom, Ubisoft, Tencent and Warner Bros. Games have said they’ll adopt it.

Does DLSS 5 increase FPS?

The neural rendering feature reduces FPS. In Engadget’s test on an RTX 5090 at 4K, NBA 2K27 dropped from about 240 to about 150 frames per second with DLSS 5 on. Upscaling and frame generation, which are also part of DLSS 5, can recover performance.

Does DLSS 5 use more power?

Yes. Engadget measured an RTX 5090 rising from about 400 watts to about 525 watts with DLSS 5 enabled. Make sure your power supply is adequate for your card.

Why are some artists and developers unhappy about DLSS 5?

Critics say it can override a game’s intended art direction, for example brightening deliberate shadows or smoothing faces, and that applying a realism-trained AI model could make different games look alike. Nvidia says developers keep artistic control and that the feature is optional.

Is DLSS 5 worth upgrading for?

Not on its own, at least not yet. Support is limited to a handful of titles, the visual gains are mostly in close-ups, and the performance and power costs are significant. If you’re buying a new card anyway, it’s a nice bonus.

What’s the difference between DLSS 4.5 and DLSS 5?

DLSS 4.5, released in January 2026, improved upscaling with a second-generation transformer model and added Dynamic Multi Frame Generation. DLSS 5 adds neural rendering, which enhances the image’s lighting and materials rather than just reconstructing it.

Does AMD have something like DLSS 5?

AMD’s FSR Redstone includes ML upscaling (FSR 4), ML frame generation, ray regeneration and radiance caching on Radeon RX 9000 series cards. AMD hasn’t announced a direct equivalent of DLSS 5’s neural rendering pass.

Can I use DLSS 5 on a laptop?

Yes, on RTX 50 series laptops and RTX Spark laptops. Expect the performance and heat impact to be felt more on thin laptops with tight power limits.

The bottom line

DLSS 5 marks a real turning point in how PC graphics work. For eight years, DLSS was about reconstructing what a game was trying to draw, faster. DLSS 5 starts adding visual detail the game never computed. That’s a genuinely new idea, and it’s no surprise it has people excited, uneasy, and sometimes both at once.

Right now, the honest picture is mixed. In the one shipping game, testers found the effect real but mostly visible in close-ups, at a significant cost in frame rate and power. The art direction debate is legitimate, and how studios use the controls Nvidia gives them will decide whether DLSS 5 becomes a respected tool or an unwelcome filter.

For buyers, our advice is simple. Don’t buy hardware for DLSS 5 alone in 2026. If you’re upgrading anyway, an RTX 50 card gives you access to it, and the rest of the DLSS suite is excellent. Make sure you have enough VRAM, and above all, don’t skimp on the power supply. A high-end graphics card running heavy AI features draws a lot of power, and a quality PSU is the cheapest insurance you can buy for very expensive hardware.

If you’re planning an upgrade and want a sanity check on whether your power supply, case and cooling can handle a new card, get in touch. We’d much rather help you get it right the first time than see a nice graphics card let down by the parts around it.

Sources

Planning a build or an upgrade?

Tell us what you already have and we'll help you spend only where it matters.

Browse the shop Ask on WhatsApp