Nvidia.com

Nvidia has officially confirmed DLSS 5 will go live Wednesday night at midnight ET — September 3 at 9 PM PT — in NBA 2K27, the sole launch title for the company's most significant graphics technology since real-time ray tracing debuted in 2018. Alongside that confirmation, Nvidia released its first official in-house performance benchmarks. Those numbers are spectacular on their face: 370 frames per second at 4K for the RTX 5090, 594 FPS at 1440p, 258 FPS for the RTX 5060 at 1080p. They are also the product of a six-to-one multiplier that Nvidia's slides disclose only in the footnotes — and stripping that multiplier out reveals a story meaningfully different from what the headlines suggest, as Tom's Hardware first reported.

What DLSS 5 Actually Does — and Why It Cannot Run on Your RTX 4090

DLSS 5, which Nvidia officially calls 3D-Guided Neural Rendering, represents a structural departure from every prior version of the technology. DLSS 1 through 4 were fundamentally performance tools: they upscaled from a lower internal rendering resolution, or synthesized additional frames between rendered frames to push the displayed frame rate higher. DLSS 5 takes a different approach entirely. After the game engine produces a conventionally rendered frame, DLSS 5 runs a locally-executed generative AI model on that frame — using its color buffer, motion vectors, and internal engine data such as surface normals and lighting buffers — to infer and enhance material properties, lighting behavior, and fine-surface detail that real-time rendering typically approximates at best, as TechTimes SIGGRAPH coverage documented.

The specific effects it targets are the ones that cost the most to render correctly: subsurface scattering on skin (in reality, only about 6% of skin's visible reflectance is direct surface reflection; the remaining 94% is light that penetrates, scatters internally, and exits elsewhere on the surface), light transmission through hair, and per-pixel contact shadows. Rather than physically simulating these effects — which can take orders of magnitude more GPU time — DLSS 5 learns from a larger offline model and runs a compact one-step pixel-space diffusion transformer that processes a single rendered frame in roughly the time a 60 FPS game has available per frame.

That diffusion transformer runs in FP8 precision — 8-bit floating-point — the inference format that Nvidia's fifth-generation Tensor Cores, present only in the Blackwell RTX 50-series, are specifically designed to accelerate. The DLSS 5 library itself, as discovered by Brazilian hardware researcher Renan Maniero inside NBA 2K27's early-access build on August 26, weighs 158 megabytes and contains approximately 148 million FP8 parameters — roughly triple the size of a typical DLSS 4 Super Resolution library, as TechTimes reported in the DLL leak analysis.

This FP8 requirement is why Nvidia gave a flat answer at a Gamescom technical briefing when asked whether older cards would ever be supported: "It's RTX 50-series only." Modders have managed to run DLSS 5 on RTX 4090 hardware with roughly 36–39% performance reduction. On RTX 30-series hardware, which lacks native FP8 acceleration, frame rates collapsed in unofficial tests. Nvidia has not announced any plan to extend official support to prior-generation architectures.

What the Official Benchmarks Actually Show — and What the Footnotes Reveal

Nvidia's first official performance figures, shared alongside the launch confirmation, present DLSS 5 in its best light. At 4K with ray tracing on Ultra settings, DLSS Super Resolution in Performance mode, and Multi Frame Generation active, the RTX 5090 posts approximately 370 FPS and the RTX 5080 reaches about 233 FPS. At 1440p with Quality Super Resolution, the RTX 5090 climbs to roughly 594 FPS, the RTX 5080 to 413 FPS, the RTX 5070 Ti to 352 FPS, and the RTX 5070 to 261 FPS. At 1080p on Ultra with the same 6X MFG configuration, Nvidia's Gamescom slides show the RTX 5090 at 797 FPS, the RTX 5080 at 602 FPS, the RTX 5070 Ti at 530 FPS, the RTX 5070 at 386 FPS, the RTX 5060 Ti at 325 FPS, and the RTX 5060 at 258 FPS.

Then check the footnotes. Every one of those numbers was captured with Multi Frame Generation 6X mode.

Multi Frame Generation in 6X mode means the GPU renders one frame and AI generates five additional frames between it, producing a displayed output that is six times the GPU's actual render rate. Dividing each headline figure by six reveals what the GPU is actually producing: the RTX 5090's 370 FPS at 4K implies roughly 62 frames per second of real rendering. The RTX 5080's 233 FPS implies about 39 FPS of actual rendering. At 1440p, the RTX 5070's 261 FPS headline implies a TechSpot-calculated render rate of roughly 44 FPS. At 1080p, the RTX 5060's 258 FPS headline implies a render rate of approximately 43 FPS.

This also helps explain why Nvidia shared 4K results only for its two most powerful desktop cards. Anything below the RTX 5080 would, at 4K, produce a render rate that looks less compelling — despite potentially still producing a high displayed frame rate through 6X frame multiplication.

A separate slide Nvidia shared makes the gap explicit. It shows NBA 2K27 running at 227 FPS natively at 4K — no DLSS, no MFG. Enable DLSS 5 with Performance Super Resolution and 6X MFG, and the displayed frame rate rises to 370 FPS, per TechSpot's benchmark analysis. What the slide does not label clearly: the 370 FPS configuration implies a render rate of roughly 62 FPS, meaning the GPU is actually rendering at about 27% of its native frame rate. The visual fidelity gain is real; the net rendering throughput is substantially lower.

Does What Renders at 44 FPS Actually Feel Like 44 FPS?

The answer requires understanding how Multi Frame Generation interacts with input latency. The five AI-generated frames that appear between each rendered frame display smoothly on screen, but they do not carry new input data — because they were generated from information available at the time the previous real frame was rendered. A displayed frame rate of 261 FPS on the RTX 5070 at 1440p, with a 44 FPS render base, responds to input like a 44 FPS game, not like a 261 FPS game.

With Nvidia Reflex enabled, Multi Frame Generation typically adds approximately 8 to 12 milliseconds of input latency relative to native rendering. That penalty is essentially fixed: it does not increase meaningfully whether the multiplier is 2X, 4X, or 6X. A game displaying 258 FPS with a 43 FPS render base still responds like 43 FPS — the additional generated frames create visual smoothness but do not create additional game-state information the display can reflect back to the player.

Nvidia's Director of Applied Deep Learning Research, Edward Liu, noted at Gamescom that "DLSS 5 is intended to be used for photorealistic situations and characters." For a sports game with crowd-filled arenas and real-athlete face captures, that framing is coherent. For competitive shooters, where input latency margin is measured in milliseconds, DLSS 5's rendering overhead and MFG latency addition represent a meaningful tradeoff that the headline frame rates do not communicate.

Is 44 FPS Baseline Enough to Run DLSS 5 Well?

The question of whether the RTX 5070's approximate 44 FPS render rate at 1440p constitutes a usable experience turns on what "good baseline" means for this technology. TechSpot's analysis concluded that the RTX 5070 "doesn't look like it will be able to use DLSS 5 at a particularly attractive render rate at 1440p" and called the RTX 5070 Ti's roughly 59 FPS render rate "borderline." The RTX 5060's approximately 43 FPS at 1080p sits squarely in sub-60 FPS territory before any frame multiplication occurs.

For context, DLSS 5 carries a performance cost of roughly 50 to 60 percent relative to running the game without Neural Rendering at otherwise identical settings — Nvidia's own stated figure, corroborated by the unofficial DLL tests that showed an RTX 5070 Ti dropping from 71 FPS to 35 FPS at 4K in Control. In the official implementation, properly integrated by Visual Concepts into NBA 2K27's engine, that penalty should be lower than the DLL-injection test because the neural rendering pass runs before upscaling rather than after it — a pipeline position that substantially reduces the pixel count the model processes. Selective per-object application, a feature the NBA 2K27 partnership specifically enabled, allows Visual Concepts to apply DLSS 5 only to character models and specific scene elements rather than every pixel in every frame — a further optimization not captured in the headline benchmarks, as TechTimes' DLL analysis showed.

Whether the visual gains justify the rendering overhead will be answered only by independent tests once DLSS 5 goes live Wednesday night. TechSpot, Tom's Hardware, Digital Foundry, and other hardware outlets will produce untethered hands-on benchmarks shortly after the midnight ET activation.

Why a Basketball Game? Visual Concepts and the Controlled-Environment Logic

Choosing a basketball simulation as the launch platform for a technology Nvidia describes as capable of Hollywood-grade photorealism drew immediate commentary, with Tom's Hardware calling it "a fairly strange" decision. The choice has specific logic. NBA 2K27 features well-lit, controlled arena environments and uses real-world athlete likenesses scanned to high fidelity — exactly the photorealistic character context Edward Liu described as DLSS 5's intended application domain.

That domain specification matters because the other context — stylized games — proved disastrous in unofficial tests. The DLL leak led to DLSS 5 applied to Final Fantasy characters (producing what GameSpot described as "horror game" results) and The Legend of Heroes characters (described as "utterly harrowing"). DLSS 5's learned model is trained on photorealistic lighting inference; apply it to a character whose appeal is its stylized departure from photorealism, and the model produces photorealism that overrides the artist's intent — a problem GameSpot documented at Gamescom.

Peter Kavic, a Senior Producer on NBA 2K27 at Visual Concepts, said the team's ability to set "the overall tone and style to match our existing art direction" was the enabling condition for the partnership. The sports-broadcast visual aesthetic NBA 2K27 has chased for a decade — maximum resemblance to a television broadcast, not stylized departure from it — makes it the most natural possible debut vehicle for a technology designed to make game characters look like they are lit by a studio cinematographer, as Tweaktown's developer interview explained.

The partnership also required that Visual Concepts control where DLSS 5 applies. The SIGGRAPH-confirmed developer toolkit — per-object engine masks, Structure Intensity and Tone Intensity sliders, three selectable AI models — enables exactly that granularity. In NBA 2K27's implementation, DLSS 5 applies to athlete character models in ways the development team controlled directly; it does not run over every pixel in every scene.

Confirmed follow-on titles include STAR WARS Zero Company, Onimusha: Way of the Sword, and The Blood of Dawnwalker. All three lean toward photorealism or stylized-but-detailed aesthetics rather than anime or classic-game art directions — suggesting Nvidia has calibrated its launch partner selection carefully.

GeForce NOW and Who Can Actually Access DLSS 5

DLSS 5 will be available at launch not only for RTX 50-series desktop and laptop GPU owners, but also for GeForce NOW subscribers at the Ultimate tier, which runs on RTX 5080-powered cloud servers. That access path matters for the millions of GeForce NOW subscribers who do not own RTX 50-series hardware — it extends DLSS 5 beyond the relatively small Blackwell installed base without requiring a GPU upgrade. Nvidia has gated cloud access to its highest-tier streaming servers specifically to ensure sufficient tensor throughput for Neural Rendering to perform as intended.

For RTX 40-series owners, the situation remains unchanged: no official support. Modders have demonstrated DLSS 5 running on RTX 4090 hardware with a 36–39% performance hit, a less severe reduction than the 45–51% measured on Blackwell in the leaked DLL tests — which is not surprising, since Ada Lovelace has some FP8 acceleration capability. Nvidia has not announced plans to extend support officially.

The Controversy Has Quieted — but Has Not Resolved

DLSS 5's debut at GTC 2026 in March drew an 84% YouTube dislike ratio, accusations of "AI slop," and a community revolt centered on the technology visibly overriding artist intent — including reports that developers at Capcom and Ubisoft learned their games had been modified at the same time the public did. Jensen Huang called the critics "completely wrong," then softened on the Lex Fridman Podcast, saying he did not love AI slop himself, as TechTimes reported at SIGGRAPH.

By Gamescom, the messaging had shifted substantially. Nvidia's presentations led with "preserve artistic intent," "artist-directed output," and "respects the rendered frame." TechSpot's analysis of the Gamescom presentation noted that Nvidia has "drastically toned down DLSS 5's output compared with the original reveal in March" — the result is subtler, less transformative, and represents "less of a slopified look, at the expense of DLSS 5 doing less." TechRadar's hands-on reporter described the NBA 2K27 implementation as "truly next-level" in terms of character lighting and shadow detail, while noting the lingering concern about whether other developers will use the technology as carefully.

Engadget's assessment was less enthusiastic: the feature "seems like it's mainly meant for pixel-peeping sickos" and "most gamers won't miss it one bit," pointing to the gap between what DLSS 5 costs in rendering overhead and the subtlety of the visual improvements in Nvidia's current presentations.

Can RTX 5060 and 5070 Owners Actually Use This at Launch?

Based on Nvidia's own numbers, the picture for mid-range Blackwell owners is mixed. At 1080p — the resolution where the RTX 5060 and 5070 will most likely be running NBA 2K27 — both cards render below 65 FPS before DLSS 5's frame multiplication: the RTX 5070 at approximately 64 FPS and the RTX 5060 at approximately 43 FPS. After 6X MFG, those become 386 FPS and 258 FPS respectively — smooth on a high-refresh display, with the latency caveat discussed above.

Whether a 43 FPS render base produces acceptable image quality in motion with DLSS 5 is precisely the question independent reviewers will investigate once the game goes live Wednesday night. DLSS 5 operates frame by frame, using motion vectors to maintain temporal stability — so the quality at 43 FPS real rendering is not the same as the quality at 43 FPS native rendering, where temporal sampling artifacts would be more visible. But it also is not the quality of 258 FPS native rendering, and Nvidia has not published a direct comparison between DLSS 5 at these render rates and native rendering at equivalent CPU settings.

For owners of the RTX 5060 and 5070 considering DLSS 5 as a factor in their GPU choice, the honest buyer's framework is: the displayed frame rate figures are real as displays go; the game-state responsiveness corresponds to the render-rate figures, not the displayed-rate figures; and the visual quality results will not be properly assessed until independent testers publish hands-on results Thursday morning.

Frequently Asked Questions

What do Nvidia's DLSS 5 benchmark figures actually mean for game performance?

Nvidia's official benchmark figures were captured with Multi Frame Generation in 6X mode — meaning one real rendered frame plus five AI-generated frames per displayed frame. To find the GPU's actual rendering output, divide the headline figure by six. The RTX 5090's 370 FPS at 4K represents approximately 62 FPS of real rendering; the RTX 5080's 233 FPS represents approximately 39 FPS. The displayed frame rate is real and reduces visual choppiness, but game-state responsiveness — how quickly your inputs appear on screen — corresponds to the underlying render rate, not the displayed rate. TechSpot's analysis of the official Nvidia benchmark figures documents this in full.

Does Multi Frame Generation make DLSS 5 unsuitable for competitive or fast-paced gaming?

Multi Frame Generation adds approximately 8 to 12 milliseconds of input latency (with Nvidia Reflex enabled), regardless of how many AI frames are being generated. Generated frames do not carry new input data: a displayed rate of 261 FPS from a 44 FPS render base still responds like 44 FPS to player input. For competitive shooters, where reaction time margins are measured in single-digit milliseconds, that latency cost is meaningful. For single-player or sports games where visual smoothness is the priority, the tradeoff may be acceptable. The NBA 2K27 implementation — a non-competitive single-player and local-multiplayer experience — represents an appropriate application context. An independent analysis of frame generation input lag mechanics explains why the penalty is fixed regardless of multiplier.

What is the actual technology inside DLSS 5, and why does it need a new GPU?

DLSS 5 uses a compact one-step pixel-space diffusion transformer — a type of generative AI architecture distilled from larger offline models — to run in real time on each rendered frame. The model's 148 million parameters are stored in FP8 (8-bit floating-point) format, the inference precision natively accelerated by the fifth-generation Tensor Cores found only in Nvidia's Blackwell RTX 50-series chips. The model processes the rendered frame's color data, motion vectors, and internal engine buffers (surface normals, lighting) to infer how lighting and material properties should look, then enriches the frame accordingly. Because FP8 inference on fifth-gen Tensor Cores is central to the model's ability to complete within a frame's time budget, there is no fallback path for prior-generation hardware. TechTimes covered the August DLL leak analysis that revealed the 148M FP8 parameters.

Will DLSS 5 ever support RTX 40-series GPUs?

When asked directly at a Gamescom technical briefing, Nvidia gave a flat answer: "It's RTX 50-series only." Modders have gotten DLSS 5 running on RTX 4090 hardware with a 36–39% performance reduction, which is less severe than on older architectures — suggesting Ada Lovelace's existing FP8 support makes it technically feasible — but Nvidia has not announced any plans for official RTX 40-series support. TechPowerUp's coverage of the official launch confirmation documents the Gamescom briefing exchange in full.