Nvidia's 260 FPS Is 43 Real Frames and a Diffusion Model

TL;DR

On September 3 at 9pm Pacific, Nvidia shipped DLSS 5 and its headline feature, 3D-Guided Neural Rendering: a pixel-space diffusion transformer that takes a finished game frame and repaints its lighting, skin, hair and materials in real time. It runs in exactly one game, NBA 2K27, on exactly one hardware generation, the GeForce RTX 50 series. Nvidia's own benchmark puts an RTX 5070 at 260 fps at 1440p, and the footnote sets Multi Frame Generation to 6X, which means roughly 43 of those frames per second come out of the game engine and the other 217 come out of a model.

What actually shipped

Everything through DLSS 4 was reconstruction. Super Resolution upscaled, Ray Reconstruction denoised, Frame Generation interpolated. Nvidia's own framing is that DLSS 4 already produces 23 of every 24 pixels you see, but all of it worked from data the engine had computed.

DLSS 5 changes the job. The model now adds detail the engine never rendered: subsurface scattering in skin, light passing through ears and hair, deeper contact shadows, material response that a real-time budget normally forces artists to strip out. Nvidia's inputs are the rendered color frame plus motion vectors, surface albedo, normals and scene semantics, so the network knows which pixels are jersey, which are skin and which are a light source.

The architecture, per Nvidia's Gamescom session as read by Igor's Lab, is a Pixel Space Diffusion Transformer operating in 2D image space rather than on scene geometry. One frame in, one frame out, deterministic, with no batch of future frames to peek at.

That constraint is the entire design. It is a colorist, not a novelist: the engine hands over a locked shot with geometry, camera and animation already decided, and the model is only allowed to change how the light lands on it. That is why it does not drift between frames or grow a sixth finger halfway down the court.

Read the footnote before the frame rate

Nvidia's numbers for NBA 2K27 look enormous. At 4K with the Ultra preset and ray tracing on, an RTX 5090 reaches up to 370 fps. At 1440p the same settings give up to 590 fps on the 5090, 410 on the 5080, 350 on the 5070 Ti and 260 on the 5070. TweakTown read the same chart and logged an RTX 5060 at 258 fps at 1080p.

At 6X, one engine frame is followed by five generated ones. So the 5070's 260 fps is about 43 rendered frames per second wearing a very good disguise, and each of those 43 was itself upscaled by Super Resolution from a lower internal resolution before the diffusion pass touched it. Nvidia is not hiding this, it is in the settings footnote, but the number on the slide is the one that travels.

None of that makes DLSS 5 fake. It does mean the honest way to compare it to a native pipeline is to compare base frame rates, not the marketing rate, and then ask whether the relit image is worth the base frame rate you gave up.

It eats more than half your frames

PCWorld puts the cost of the neural rendering pass on its own at 50 to 60 percent of frame rate. Super Resolution and frame generation exist to pay that bill, which is why the two ship as a bundle and why the benchmark footnote reaches for 6X.

Some context on how tight the budget is. When Nvidia revealed the technique at GTC in March, the demo needed two RTX 5090s. In its launch post the company writes that it went "from running on dual GeForce RTX 5090s to now a single graphics card, we've achieved a 5X performance gain in just six months." That is a real engineering result. It also tells you how far outside the real-time envelope this started.

The artists got sliders

The obvious fear with a model that repaints your game is that it repaints your game. Nvidia's answer is developer-side control: Structure Intensity, Tone Intensity, semantic masks so faces, cloth and background are not processed identically, and selectable model variants. Visual Concepts used them to keep the facial geometry scanned from real players intact while letting light pass through ears and hair.

The early verdict is that it works, and that it exposes the next weakest link. PCWorld found the on-court footage close to broadcast, then watched the same players turn into "robotic Pinocchios" in cutscenes, because photoreal lighting on 2020s animation makes the animation the thing you notice. Engadget was blunter, calling DLSS 5 something "mainly meant for pixel-peeping sickos."

The audience matters here, because the people who make games are not in a generous mood about generative AI.

Those figures come from GDC's 2026 State of the Game Industry survey, and they are the reason Nvidia spent so much of its launch material on the word control. DLSS 5 is not a text prompt replacing an artist's work, it is a pass that happens after the artist's work, on the player's machine, at a strength the studio picks. Whether that distinction survives contact with a comment section is a separate question.

What Nvidia has not published

For anyone trying to reason about this as an inference workload rather than a marketing beat, the interesting gaps are large. Igor's Lab notes there is no published model size, no VRAM figure and no per-frame inference time, and no demonstration of temporal stability under fast motion. Nor is there any evidence yet of how the model behaves on a stylized game, where "more photoreal" is an active hostile act against the art direction.

The scope is small on purpose. RTX 50 series desktop and laptop only, plus GeForce NOW Ultimate on RTX 5080-class rigs, and a launch library of exactly one basketball simulator. Nvidia's own Edward Liu set the ceiling for everyone: "We're still really orders and orders of magnitude away from really simulating the real world in real time."

Key Takeaways

• DLSS 5 shipped September 3 in one game. NBA 2K27, RTX 50 series only, plus GeForce NOW Ultimate. Everything else is a promise.
• It is a relighting pass, not an upscaler. A pixel-space diffusion transformer takes the engine's finished frame plus motion vectors, albedo, normals and semantics, and returns one deterministic frame with material and lighting detail the engine never rendered.
• Divide the benchmark by six. Nvidia's fps figures use Multi Frame Generation at 6X, so the RTX 5070's 260 fps at 1440p is roughly 43 engine-rendered frames per second, each already upscaled from a lower internal resolution.
• The pass is expensive. PCWorld measures a 50 to 60 percent frame rate cost for neural rendering alone, and Nvidia needed a claimed 5X optimization in six months just to get it off dual RTX 5090s.
• Artist control is the whole pitch. Structure Intensity, Tone Intensity, semantic masking and model variants exist because 52% of surveyed game developers already think generative AI is hurting their industry.
• The unpublished numbers are the ones to ask for. No model size, no VRAM cost, no per-frame latency, no stylized-game evidence. Until those land, treat the visual comparisons as a demo, not a spec.

Sources: Nvidia GeForce News: DLSS 5 3D-Guided Neural Rendering, Igor's Lab, TweakTown, PCWorld, Engadget, The FPS Review, GDC State of the Game Industry 2026