DLSS 4 Explained: RTX 50 Series’ Performance Multiplier

BossMac Suba
By
BossMac Suba
Boss Mac Suba is the driving force behind Back2Gaming.com, one of the most enduring, dedicated PC gaming hardware and video game reviews publication. With over a...

GeForce RTX 50 Series Arrives

During the announcement of the GeForce RTX 50 series graphics card at CES 2025, the biggest highlight for the launch segment was this slide:

An RTX 5070 will be enough to compete with the RTX 4090 at $549. The internet was immediately in chaos with polarizing opinions and there’s this tired gem “I’ll pay for your new RTX 5070 to replace your “slow” RTX 4090.” Regardless on where you stand, NVIDIA has been using this marketing since the debut of RTX 20 series and depending how you treat it is usually where you stand.

Now there’s no published data on how many users actually use DLSS and my personal experience always tells me to go for higher framerates in most games. DLSS is a decent compromise in increasing performance while maintaining details and if you’re a visual learner, you can jump straight to3:49 of the video below:

With DLSS 4, NVIDIA steps this up further not as they not only upscale, they’re now also not just improving the existing Frame Generation technology introduced in DLSS3 but multiplying it further with DLSS4 Multi-Frame Generation.

Deep-Learning Super Sampling Revisited

Deep Learning Super Sampling is NVIDIA’s technology that uses AI upscaling to create faster frame rate. In native raster rendering, a gaming running at 60FPS at 1920×1080 will most likely perform faster if the game resolution is set to 1280×720. This is the principle at which all DLSS release version operate in: to run the game at a lower resolution proportional to the actual game resolution and then use parallel AI calculation to fill in the pixels required to complete the image.

And I mentioned two things to note in that last sentence: DLSS versions and parallel AI calculation, or to be specific, parallel compute. Let’s talk about parallel computation first:

Parallel Processing: CUDA, RT, Tensor Cores

CUDA cores do the primary heavy lifting in the traditional NVIDIA render pipeline. This is what most people call the “raw” or “brute” rendering engine of NVIDIA graphics cards. In the last decade though, CUDA cores have found a new calling in various applications most notable cryptocurrency mining and AI/ML. While mining has faded, AI is nowhere near slowing down at the start of 2025.

GeForce RTX graphics cards are now being offered as AI engines for gamers, power users, and professionals and while use-cases still vary, it can’t be denied the industry has shifted in this direction. With the transition of GeForce from GTX to RTX, we are introduced to RT cores and Tensor cores. RT cores are fixed-function hardware designed to handle the complex task of computing light rays and how it behaves in 3D space. In modern cinematics, CG movies are done using path-tracing while in games, ray-tracing is currently the norm.

We have seen games like Cyberpunk 2077 and Minecraft introduce path-tracing which shows how realistic path-traced lighting can make in-game objects. But it is very slow. EXTREMELY SLOW. To be really honest, we’re are far from even making 1080p path-traced games playable.

To offset this, NVIDIA uses Tensor cores to handle the AI/ML computation for upscaling, image reconstruction, and frame generation. This means that for every frame rendered, all 3 of these core types are doing work producing the final frames displayed on-screen. For the majority of applications, the final benchmark for this how much frames-on-screen can we see.

I personally am very receptive of how AI is positioned but I am very specific when it comes to how it’s marketed and NVIDIA, isn’t really perfect in this respect either.

Which brings us back to DLSS. This process is specific to the Tensor core functions and progress on this technology is quite rapid as the RTX 50 series sees the debut of DLSS4. And a note on DLSS versions, they are release steppings rather than versions of a single technology. The primary steppings are as follows:

  • DLSS – the initial upscaling model featuring a direct upscaling model
  • DLSS2  – an improved upscaling model featuring finer controls and improvements
  • DLSS3 – introduces Frame Generation which takes 2 frames and inserts a completely AI-generated frame in-between. Limited to RTX 40 cards.
  • DLSS3.5 – adds Ray Reconstruction which improves upscaled details by recreating details normally diluted by the upscaling process
  • DLSS4 – introduces Multi-Frame Generation which inserts up to 3 frames, uses less VRAM, and requires only a single frame. Limited to RTX 50 cards.

Games and applications developers can choose to adopt any of the versions but older games that are not maintained would probably see no updates in what DLSS implementation it uses.

DLSS4

DLSS4 is a multi-pronged update to NVIDIA’s AI-based rendering technologies, introduced alongside the new GeForce RTX 50 Series. At the foundation, DLSS4 retains the upscaling and frame generation approach of DLSS3 and DLSS3.5, but it now leverages new hardware and software features for improved performance and image quality.

  1. Multi-Frame Generation
    • Key Concept: Traditional Frame Generation (as seen in DLSS3) analyzes two sequentially rendered frames (plus motion vectors and optical flow) to create an interpolated in-between frame. DLSS4’s Multi-Frame Generation (MFG) goes further by creating multiple AI-generated frames between each traditionally rendered frame.
    • Hardware Requirements: Generating additional frames in between requires heavy AI inference. The RTX 50 Series incorporates a revised optical flow and dedicated hardware changes (fixed-function Tensor cores under Blackwell architecture) that enable more precise optical flow calculations and improved frame pacing. NVIDIA has noted before it is possible to run newer models in older hardware but the performance penalty doesn’t really provide anything to the overall experience. With the propensity of the tech media to tear into anything, I’d be reserved in enabling it myself.
    • Frame Pacing: To manage multiple AI-generated frames, NVIDIA is using a technique referred to as flip metering in the display engine. This helps avoid irregular pacing or “micro-stuttering” that can happen if frames get queued unevenly.
  2. Enhanced Single-Frame Generation

    • For RTX 40 & 50 Series: A separate AI path aims to boost performance and reduce memory consumption during single-frame generation. While Multi-Frame Generation is only available on RTX 50 hardware, owners of RTX 40 Series GPUs can still benefit from an improved frame generation model that is more optimized in terms of VRAM usage.
  3. Transformer-Based Upscaling & Ray Reconstruction

    • Shifting Beyond CNNs: Earlier DLSS versions (1.x through 3.5) used convolutional neural networks (CNNs) for upscaling and image reconstruction. DLSS4 introduces Transformer-based models that process a wider spatial and temporal context—effectively “seeing” more of the image across both space and time.
    • Super Resolution and DLAA: The new Transformer approach applies to both the Super Resolution pipeline (which upscales from a lower render resolution) and DLAA (which is essentially the same logic without an upscaling factor, focusing on anti-aliasing quality). The net effect is improved detail retention and more stable edges across consecutive frames.
    • Ray Reconstruction: Path- and ray-traced content tend to produce a lot of “noise” that must be denoised or reconstructed. In DLSS3.5, Ray Reconstruction was introduced to mitigate these artifacts; DLSS4 continues refining that approach with improved denoising logic derived from Transformer models.
  4. Performance Implications

    • Workload Distribution: With DLSS4, each pixel in a final frame is increasingly likely to be generated or enhanced by AI. According to NVIDIA’s own breakdown, Multi-Frame Generation pushes this ratio even higher compared to prior versions.
    • VRAM and Execution Efficiency: The new Tensor Core design and refined optical flow calculations reduce overhead on the GPU. This means VRAM usage is optimized, so real-time upscaling and multiple AI-generated frames place less strain on the memory subsystem compared to older DLSS releases.
  5. Integration & Game Support
    • Driver-Level Overrides: NVIDIA provides an interface (referred to as the “NVIDIA App” in your draft) where users can enable or tweak DLSS settings, including Multi-Frame Generation—assuming the game already supports Frame Generation. This approach decouples some of the overhead from the game engine, simplifying integration for developers.
    • Backwards Compatibility: Existing DLSS3-enabled titles can often be updated to harness the new Multi-Frame Generation feature, but the level of effort may vary. Games that do not support Frame Generation at all will not automatically see these benefits without an update.
  6. Path-Tracing & Future Outlook
    • Full Path-Traced Titles: With more aggressive AI-driven rendering, NVIDIA is laying groundwork for heavier path-traced workloads. Even on RTX 50 Series hardware, path tracing can still be resource-intensive, but advanced upscaling and multi-frame interpolation can reduce perceived performance hits.
    • Transformer Growth: The pivot to Transformers indicates a longer-term strategy. Transformers scale well with more data; as GPUs evolve, we can expect more real-time AI-based operations that further close the gap between real-time and cinematic rendering.

Addressing Input Latency with Reflex 2

Reflex 2 is NVIDIA’s latest refinement to its input-latency reduction toolkit, succeeding the original Reflex introduced alongside the RTX 30 Series. In scenarios where AI-based Frame Generation (including multi-frame generation) might artificially inflate perceived lag—because the GPU is generating additional frames in between traditionally rendered frames—Reflex 2 helps synchronize the game engine, CPU, and GPU work more efficiently. By doing so, it mitigates or reduces any added latency from the extra frames.

  • Frame Queue Management
    Reflex 2 dynamically meters how rendered frames are queued on the GPU. This ensures game logic and rendered output stay closely aligned, rather than buffering multiple frames, which could otherwise contribute to sluggish input responses.
  • Tighter Integration with AI Pipelines
    Since Multi-Frame Generation (and even single-frame generation) uses advanced optical flow and motion vector data, Reflex 2 works in tandem by prioritizing the immediate inputs from the mouse and keyboard at the most crucial stage in the rendering pipeline. The result: the AI has more up-to-date motion information, further minimizing incongruous or “laggy” frames.
  • Practical Outcome
    With Reflex 2 enabled, a user running DLSS4 can potentially see high frames per second and maintain responsiveness closer to “native” levels. It won’t eliminate the overhead from generating frames completely, but it narrows that gap compared to having no latency optimization in place.

Ultimately, DLSS4 is intended to ensure that while baseline raster performance may reach saturation points at higher resolutions, AI-based approaches can still push effective frame rates and fidelity upward. By combining more sophisticated upscaling, multiple interpolated frames, and better denoising, DLSS4 sets the tone for the RTX 50 Series’ hallmark feature: using specialized cores and advanced AI models to squeeze more output from every rendered pixel.

Closing Thoughts: AI Frames and the Road Ahead

The debate over so-called “fake frames” remains a central talking point for AI-based rendering. Critics argue that frames generated by inference aren’t “real,” yet the visual output is genuinely displayed and consumed by the player. This highlights a broader shift in the graphics pipeline, wherein traditional brute-force rendering is making way forAI-accelerated techniques e.g. DLSS4.

A key concern/complaint tied to AI-generated frames is added latency. While Multi-Frame Generation (MFG) can boost perceived frame rates, each extra frame may compound timing issues between input and on-screen response. Reflex 2—the latest iteration of NVIDIA’s latency reduction technology—is designed to counter this effect. It coordinates CPU-GPU workloads more precisely to reduce frame queuing and ensure the AI has timely motion data to work with. The result is smoother motion and tighter synchronization, mitigating some of the “lag” that might otherwise arise from interpolating multiple frames.

These developments indicate a larger computational paradigm shift. As game engines gradually move beyond approximate ray tracing and inch closer to full path tracing, they will become ever more reliant on AI to handle upscaling, denoising, and frame interpolation duties. The line between “real” and “AI-generated” pixels will continue to blur, reflected not just in DLSS4’s multi-frame approach but in how latency-reduction tools adapt to new rendering paradigms. Whether these frames are labeled as fake or not, the goal remains the same: higher fidelity, smoother motion, and a more responsive experience for players— the direction of gaming in the near future is one in which AI computing is firmly embedded in every stage of real-time graphics.

NVIDIA invented the GPU after all, they can sure as hell remake it, especially in their current market position. With PlayStation following suit in their own design philosophy, we see both market leaders agreeing on one thing and that’s upscaling and frame generation are definitely a solution to a problem. With Playstation having stated they acknowledge that their playerbase prefer to play in High-Framerate Mode rather than in 4K Fidelity mode, this resonates with PC gamers who most likely swing into that direction as well.

Don’t get me started on the lazy devs argument either. While indie games could show a developers’ relative infancy in their skills, larger studios are under investor and publisher pressure which sets unrealistic goals for teams. This leads to compromises that if prove game-breaking, could ruin a studio altogether. So in this aspect, I am with developers taking time with their game but also urging them to pushback against publisher pressure.

Anyway, to wrap it all up. DLSS 4 is amazing from a technical standpoint and is a breakthrough in its application. But all that rests on how it works in real life. Will Reflex 2’s hard framecapping be enough to sync input to the amount of frame that DLSS4 Multi Frame Generation outputs? You’ll need to try it for yourself and as of this writing, the GeForce RTX 50 series should be in-stores soon.

Addendum

There will be 75 games expected to coincide or at least be ready with DLSS4 once the RTX 50 series arrives on shelves:

DLSS 4: Day 0 Support For 75 Games & Apps
A Quiet Place: The Road AheadGod of War RagnarökReady or Not
AkimbotGray Zone WarfareRemnant II®
Alan Wake 2GROUND BRANCHSatisfactory
Aunt FatimaHITMAN World of AssassinationSCUM
Backrooms: Escape TogetherHogwarts LegacySenua’s Saga: Hellblade II
Bears in SpaceICARUSSILENT HILL 2
BellwrightImmortals of Aveum™Skye: The Misty Isle
Crown SimulatorIndiana Jones and the Great Circle™Slender: The Arrival
Cyberpunk 2077JusantSquad
D5 RenderJX Online 3S.T.A.L.K.E.R. 2: Heart of Chornobyl
Deceit 2KristalaStar Wars Outlaws™
Deep Rock GalacticLayers of FearStar Wars Jedi: Survivor™
Deliver Us MarsLiminalcoreStarship Troopers: Extermination
DESORDRE: A Puzzle AdventureLords of the FallenStill Wakes The Deep
Desynced: Autonomous Colony SimulatorMarvel RivalsSupermoves
Diablo IVMicrosoft Flight SimulatorTest Drive Unlimited Solar Crown
Direct ContactMicrosoft Flight Simulator 2024The Axis Unseen
Dragon Age™: The VeilguardMortal Online 2The Black Pool
DugeonborneNARAKA: BLADEPOINTTHE FINALS
DYNASTY WARRIORS: ORIGINSNeed for Speed UnboundThe First Descendant
EnlistedOnce HumanThe Thaumaturge
Flintlock: The Siege of DawnOutpost: Infinity SiegeTorque Drift 2
Fort SolisPax DeiTribes 3: Rivals
Frostpunk 2PAYDAY 3Witchfire
Ghostrunner 2QANGAWorld of Jade Dynasty
TAGGED:
Share This Article
Follow:
Boss Mac Suba is the driving force behind Back2Gaming.com, one of the most enduring, dedicated PC gaming hardware and video game reviews publication. With over a decade of experience in IT and more in doing reviews for things he love since 2002 from Jpop albums, anime series, games in the early days of the internet.My stance on AI-generated writing and content

Tondo since birth. To this day.

12 Comments
    • Playstation has it na rin. Once Xbox does frame generation, magiging standard way of rendering na rin sya, di nga lang geoemtric render. Garbage-in/garbage-out naman DLSS. If its a bad game, ITS A BAD GAME haha

  • DLSS is icing on the cake and those people saying “fake frame” probably didnt experience it. Ang tagal na nang dlss the thing is i cant tell the difference if its fake or not while gaming ang importante sakin how smooth the game is

      • Ellis T***u i hope so but whats the main reason for mediocre performance uplift every gen and continuos price hike? Poor optimization? Lazy program developer? Are they doing this on purpose for business? We dont know, it is what it is. Kung ako lang madami talaga ko reklamo sa Nvidia but its our choice wether we gonna buy it or not di nila tyo pipilitin to buy their products

      • Leo A****r C******l problem is competition is way too one sided. Its a good thing we got the battlemage from intel kaya nadelay ang announcement ng 60 tier card ni nvidia. Competition is good but nvidia is just too far ahead

      • xx60 cards get announced 2-3 months after the initial launch since NV needs to build up their supplies from cutdown silicon.

Leave a Reply

Your email address will not be published. Required fields are marked *