DLSS5 STUDIO
EN
Open the Studio
GPU SUPPORT · SEPTEMBER 2026

DLSS 5 without an RTX

DLSS 5 Neural Rendering needs FP8 tensor hardware, and that one requirement draws a hard line straight through NVIDIA's own line-up. This page is where the line falls, what the community tools do on the wrong side of it, and what they cost in frames.

The short answer

Officially, today: RTX 50 and nothing else. NVIDIA has confirmed RTX 40 is coming and has given no date. Nothing older than RTX 40 has the FP8 tensor hardware the model runs on, and no AMD, Intel or Apple GPU is supported at all. Mods exist for RTX 20, RTX 30 and for RDNA 4 Radeons: they start, they take roughly half your frame rate or worse, and they are built on an experimental runtime that was never meant to leave NVIDIA's own testing.

Checked 11 September 2026

What runs it, card by card

HardwareDLSS 5Why
GeForce RTX 50 · BlackwellYes, officiallyShipped on 3 September 2026, NBA 2K27 first, with the rest of the announced list arriving through the autumn. RTX 5090 down to RTX 5050, desktop and laptop.
GeForce RTX 40 · Ada LovelaceConfirmed, no dateNVIDIA says support is coming once the model is tuned on Blackwell. Fourth-generation Tensor Cores already do FP8, so the hardware is not what is holding it up.
GeForce RTX 30 · AmpereMods onlyThird-generation Tensor Cores have no FP8 path. The swapper tools will start the runtime; the arithmetic has to happen in a wider format, and the frame rate pays for it.
GeForce RTX 20 · TuringMods onlyThe same missing format, one generation further back, and the cost grows with the distance.
GeForce GTX 10 and olderNoNo Tensor Cores at all. There is nothing on the die to run a neural model on, and no tool changes that.
Radeon RX 9000 · RDNA 4Mods onlyRDNA 4 has FP8 in hardware, which is the only reason a port exists. It is early alpha and it costs about two thirds of the frame rate.
Radeon RX 7000 · RDNA 3UnconfirmedThe mod's author says it should work and is asking for reports; nobody has published a working configuration. Everything older than RDNA 3 has no FP8 at all, and is a no.
Intel Arc · Apple silicon · integratedNoNo NVIDIA runtime and no port. Here the answer is not «not yet».

What NVIDIA actually ships today

Neural Rendering went out on 3 September 2026 as a Blackwell feature: RTX 50 desktop and laptop cards, starting in NBA 2K27, with the announced list — Resident Evil Requiem, Starfield, Hogwarts Legacy, Assassin's Creed Shadows, Delta Force and a dozen more — arriving through the autumn.

RTX 40 is confirmed and undated. NVIDIA's position is that it wants the model optimised on the newest generation first; Ada owners have an assurance and no calendar, which is a better place to be than every other card on this page but is still not a date.

There is also no override. NVIDIA confirmed the NVIDIA App will not carry a DLSS 5 toggle for games that do not implement it — no equivalent of the DLSS override that exists for earlier versions. That gap is precisely what every tool further down this page was built to fill.

Why the line falls exactly there

The DLSS 5 model runs in FP8: eight-bit floating point. Blackwell does FP8, and so does Ada — fourth-generation Tensor Cores added the format two years before there was a consumer feature that needed it. Ampere's third-generation cores do not have it. Turing's second-generation cores do not either.

That is the entire explanation for the support matrix, and it is also why RTX 40 is a promise rather than a research project: the hardware requirement is already met, so what is left is tuning. For RTX 30 and RTX 20 nothing is left to tune — the arithmetic has to be carried out in a wider format than the model was built for, and the difference comes straight out of the frame rate.

GTX cards are a shorter conversation. They have no Tensor Cores, so there is no unit to emulate on and no version of this that is a settings change.

The unofficial routes, and what they cost

Two families of tool are circulating. DLSS 5 Swapper drops NVIDIA's RTX-50-only check and will start the runtime on RTX 20 through 50. Its Feeder mode goes further: paired with ReShade — which supplies the depth buffer and the motion vectors DLSS needs and the game never exposed — it injects DLSS 5 into DirectX 9, 10, 11 and 12 titles that have no DLSS of any kind, and into emulators including PCSX2, Dolphin, Duckstation and Cemu.

The other route is ReShade with RenoDX directly, which is where the experiment started. A modder swapped in Ada-compatible binaries and had it running on RTX 40 cards before NVIDIA shipped anything for them.

The cost is not subtle. On an RTX 5070 Ti — a card that officially supports the feature — modding DLSS 5 into Control took it from 71 fps to 35. That is roughly half, on hardware that is not emulating anything. Move to a card without FP8 and the number gets worse, not better.

And the binary at the centre of all of it is an unoptimised experimental build that escaped NVIDIA's own testing. Whatever you measure with it is a preview of the feature, not a benchmark of it — which matters most for exactly the people these tools attract, the ones trying to decide whether the feature is worth a new card.

If the card is a Radeon

There is now one path, and it is narrower than the NVIDIA ones: RDNA 4 only, through a mod that carries NVIDIA's own DLL and runs it on AMD hardware because RDNA 4 happens to have the FP8 format the model needs.

It works, in the sense that the picture comes out right. On an RX 9070 XT in Cyberpunk 2077 at 1080p it costs about two thirds of the frame rate, and online games are unavailable because anti-cheat blocks the DLL. It has a page of its own, with the numbers.

What you are agreeing to

Every one of these routes needs nvngx_dlss DLL files that ship inside games, moved into games that did not ship them. That is a licensing question, and it is yours to answer.

Anti-cheat is the practical one. An unsigned DLL placed next to a game's executable is exactly the pattern kernel anti-cheat exists to refuse, so anything with an online component is off the table — not because it performs badly, but because it will not launch or will not let you play.

And the honest framing: none of this is the feature NVIDIA will eventually ship to your card. It is the feature's engine, started outside the car, by people who are very good at starting engines.

The other way to see it

If what you want is to know what Neural Rendering does to a picture — not to run it on your own machine tonight — you can skip everything above. This site keeps RTX 40 and 50 cards powered on and hands one to you for 180 seconds: you upload a clip, the NVIDIA runtime processes it on our hardware, and you download the result together with the render report. Nothing installed, no account, no DLL moved anywhere. It is free and non-profit, and there is sometimes a queue, because the cards are real machines in a real room.

Try it without an RTX →
Take a slot

Questions this page gets

Will DLSS 5 ever come to RTX 30?

Nothing has been announced, and the hardware argues against it: Ampere's Tensor Cores have no FP8 path, so it would not be a driver update but an emulation — which is what the mods already do, at the frame cost you would expect.

Can I switch DLSS 5 on from the NVIDIA App?

No. NVIDIA confirmed there is no DLSS 5 override in the App, for any card. A game either implements it or it does not, and everything in between is a community tool.

Is a GTX 1080 Ti really out, even with a mod?

Yes, and not by a narrow margin. Neural Rendering is a model executed on Tensor Cores, and Pascal has none. A mod can move a DLL; it cannot add silicon.

Can the swapper get me banned?

In a single-player game nobody is looking. In anything with kernel anti-cheat you are placing an unsigned DLL beside the executable, which is the precise shape of the thing it exists to block. Treat online titles as unavailable rather than risky.

Are ON/OFF comparisons made with mods representative?

For image quality, broadly yes. For performance, no: the build in circulation is an unoptimised test binary, and the mod feeds the runtime depth and motion data the game was never designed to hand over.