Nvidia Queues DLSS 5 for RTX 40 After 50-Series Tune

Nvidia will bring DLSS 5 to GeForce RTX 40 cards only after it trims a frame-rate tax that still costs an RTX 5090 51.2% at 4K. The neural renderer shipped in NBA 2K27 on September 3, 2026 for RTX 50 GPUs, laptops, and GeForce NOW, and Ada owners were told to wait.

That wait is not a courtesy port. It is what you do when the current flagship still cannot run the new pass cheaply, and last-gen hardware would make the same pass look worse.

The 40-Series Promise Comes With a Queue

On the same day NBA 2K27 went live at 9PM Pacific, Nvidia staff updated the GeForce Forums FAQ. The company called DLSS 5 its most computationally demanding model to date, said model updates are due later this fall, and only then described expanding official support to RTX 40 Series.

Our current focus is on optimizing performance for the GeForce RTX 50 Series with model updates expected later this fall. Once RTX 50 Series performance is more fully tuned, we plan to work on expanding official support to the GeForce RTX 40 Series.

NVIDIA staff, GeForce Forums, September 3, 2026

Nvidia spokesperson Rajal Maharaj put the same order in plainer words. The focus is the 50-series, and the company will look at the 40-series once it feels the new cards are in a good place. There is no date. There is also no pledge for RTX 30 or RTX 20 cards, even though unofficial builds have already been forced onto those generations.

The in-game toggle is named DLSS Neural Rendering. It is optional. Visual Concepts used it in NBA 2K27 to push subsurface scatter through skin and tighter contact shadows across jerseys, hair, and crowd models, while keeping the facial scans of real players intact. On a tuned studio pass the change can be quiet. The cost is not.

A 5090 Still Loses 51% at 4K

Digital Foundry ran NBA 2K27 across the RTX 50 stack with the neural pass on and off. Founder Richard Leadbetter called the model “super, super heavy,” and the 4K numbers on the fastest GeForce card in the line are the ones that explain the Ada queue.

NBA 2K27 FRAME-RATE COST ON RTX 50 CARDS

On the 5090 that 4K pass adds about 8.3 ms of frame time. Leadbetter called eight milliseconds at 4K on a 5090 “mega expensive.” Multi Frame Generation can multiply the displayed count after that hit, but it cannot put back the base frame the neural pass just spent. That is the bind for Ada. If Blackwell still pays this bill, fourth-gen Tensor hardware will not look like a gift when the same model lands there.

Nvidia’s 370 FPS Number Stacks Extra Frames

Nvidia’s own Game Ready Driver 616.64 write-up says a GeForce RTX 5090 can hit 370 fps in NBA 2K27 at 4K Ultra with ray tracing when DLSS is on. That figure is the full stack: 3D-guided neural rendering plus Super Resolution plus Multi Frame Generation. It is not the rate of the neural pass by itself.

The launch post also walks through a 5X performance gain in six months, from dual GeForce RTX 5090s in the March demo to a single card in September. That 5X is model throughput against the March build. It is not a 5X rise in in-game frames, and it does not cancel the table above. GeForce NOW Ultimate streams the same path from RTX 5080 rigs in the cloud, which is another way of saying Nvidia wants the feature on current silicon first, including silicon it owns in the rack.

Nvidia’s ADLR lab describes DLSS 5 as a one-step pixel-space diffusion model that writes the final displayed look rather than reconstructing a more expensive reference frame. That is a different job from Super Resolution, and it is why a technology family famous for raising fps now spends fps to buy lighting and materials.

Why Official Ada Support Waits on Blackwell

The FAQ’s sequence is the policy. Tune the 50-series until the tax shrinks, then port. Maharaj’s version is the same sequence with less polish: look at Ada after Blackwell feels like a good place. A good place is not 62.7 fps at 4K on a 5090 with the new pass enabled, and it is not a 5080 that falls just over 40 fps in the same game at 4K with ray tracing.

THE DLSS 5 CLOCK FROM MARCH TO FALL

1. March 16, 2026: Nvidia announces DLSS 5 as a fall neural renderer and demos it on dual GeForce RTX 5090s.
2. August 28, 2026: A leaked nvngx_dlssnr.dll from the NBA 2K27 early-access build is patched to run on Ada Lovelace, including the RTX 4090 and RTX 4080.
3. September 3, 2026: NBA 2K27 ships DLSS Neural Rendering for RTX 50 GPUs, RTX 50 laptops, and GeForce NOW, and the forums FAQ promises official RTX 40 work after more 50-series tuning.
4. September 8, 2026: Digital Foundry’s NBA 2K27 benches land, with the 5090 4K Ultra drop at 51.2% and an 8.3 ms add-on.
5. Later this fall: Nvidia says further model updates are in active development. Official Ada support still has no date.

The second-order read is simple. Shipping Ada now would publish the same heavy model on weaker Tensor hardware and make the 50-series exclusive window look like a software lock rather than a speed floor. Holding Ada until the fall tune lands lets Nvidia move the floor first.

Uncle Burrito’s Patch Already Hits Ada Cards

The hardware wall was already thinner than the launch messaging. RTX Remix modder Uncle Burrito replaced Blackwell-only CUDA binaries inside the leaked library and ran the same neural-rendering DLL on Ada. His 4090 clip falls from around 135 fps to 82 fps with the pass on, a 39% cut, in the same band as several 50-series unofficial tests from the same week.

That 39% is not an official NBA 2K27 result, and the August build was unfinished. It is still the existence proof. Ada can execute the model. Fourth-generation Tensor Cores already speak the FP8 path this class of network uses. The 50-series card on Nvidia’s spec sheet carries fifth-generation Tensor Cores rated at 3,352 AI TOPS, and Puget Systems counted 680 fifth-generation Tensor Cores on the 5090 against 512 fourth-generation cores on the RTX 4090. Blackwell also adds FP4. Those are real gaps. They are not the same thing as “Ada cannot run it.”

By September 8, unofficial packs were already bolting DLSS 5 and multi-frame generation onto RTX 20, 30, and 40 cards, including hybrid paths that lean on AMD’s FSR 3.1, while Nvidia still keeps official Multi Frame Generation on Blackwell. The company has also pushed app-side changes that try to close those unlocks. The unofficial scene is not waiting for the FAQ’s second sentence.

The In-Game Switch Hides the Studio Dials

Maharaj said gamers get a simple on-or-off toggle. The granular artistic controls stay with developers. That split matters as much as the fps table, because the March demos and a lot of unofficial clips went ugly when people cranked intensity on faces, fabrics, and whole frames.

CONTROLS STUDIOS GET AND PLAYERS DO NOT

• Model selection: Distinct model weights with different training parameters that steer the final look.
• Structure intensity: High-frequency adds such as ambient occlusion, contact shadows, and subsurface scattering.
• Tone intensity: Low-frequency lighting and color response.
• Automatic masking: Object isolation so the pass can hit characters without hand tagging every mesh.
• Developer masking: Per-pixel custom masks so a studio can spare a hero face, a crowd, or a prop.

High-intensity unofficial passes in other games have already pushed character skins toward a beauty-filter look that no longer matches the authored art. Fortnite clips circulating after the leak made that plain when the effect was turned up. NBA 2K27’s official pass stays quieter because Visual Concepts kept those dials. Nvidia does not want the same sliders in player menus on 50-series cards, and it will not want them on 40-series cards either. Ada owners who eventually get the official path should expect the same blunt switch, not the RenoDX-style control sheet.

Fall Updates Land Before Any 40-Series Date

Nvidia says the optimization curve is not flattening and that more model updates are due later this fall. That is the work Ada is queued behind. If those updates cut the 8.3 ms 4K add-on on a 5090, a 4080 or 4090 becomes a less embarrassing second customer. If they do not, official 40-series support still has a way to arrive as a checkbox that few people leave on.

The 40-series installed base is the group with skin in this delay. Those cards already run Super Resolution, Frame Generation, and Ray Reconstruction. They already ran a patched DLSS 5 library in August. What they do not have is Nvidia’s blessing, a date, or a version of the model that has been pulled down to Ada’s Tensor budget on purpose.

Maharaj said that once the 50-series is in a good place, Nvidia is going to look at the 40-series. As of Digital Foundry’s September 8 benches, that place is still a 51.2% 4K cut on the 5090, and Ada is still in line behind it.