Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

DLSS 5 Now Runs on RTX 20, 30 and 40 Cards: The Requirement Collapsed in a Week

2026-08-31 · News and analysis

When the DLSS 5 neural rendering library leaked out of an NBA 2K27 build, every working demonstration ran on an RTX 50 series card. We wrote that up as a hardware requirement, with the reasoning that neural rendering leans on Blackwell's FP8 throughput.

That held for about four days.

Modders have since had it running on RTX 40, then RTX 30, then RTX 20 series cards. The Blackwell exclusivity was never a hardware lock. It was just where everyone started. We have gone back and corrected our system requirements article accordingly, and this piece covers what actually happened.

What Actually Changed

The first attempts on older cards went badly, and that is the interesting part.

The leaked model is built around FP8, an eight bit floating point format that Blackwell tensor cores handle natively. RTX 30 series Ampere cards do not have FP8 units. Neither do RTX 20 series Turing cards. Running an FP8 workload on hardware without FP8 units means emulating it, and the early RTX 30 results reflected that. Performance did not merely drop. It fell apart.

Then someone switched the precision. Running the model in FP16 instead, a format Turing and Ampere do support in hardware, produced dramatically better numbers on the same cards. Not RTX 50 numbers, but usable ones.

So the story is not that a lock was picked. It is that the first people to try older cards used the wrong numeric format, got terrible results, and the community found a better path within days.

The Current Compatibility Picture

GPU generation	Status	Practical notes
RTX 50 (Blackwell)	Best case	Native FP8. All the published benchmark numbers come from here
RTX 40 (Ada)	Working	The first generation modders unlocked after Blackwell
RTX 30 (Ampere)	Working via FP16	Terrible on the FP8 path, substantially better on FP16. No FP8 or FP4 units
RTX 20 (Turing)	Working via FP16	Same situation as Ampere, on older and slower tensor cores
GTX 10 and 16 series	No	No tensor cores. Some FSR based paths exist through OptiScaler, but that is not DLSS 5
AMD Radeon	No	DLSS is NVIDIA only
Intel Arc	No	DLSS is NVIDIA only

The pattern to take away is that neural rendering needs tensor cores, and every RTX generation has them. What separates the generations is how fast they are and which numeric formats they handle natively, which turns into a performance difference rather than a yes or no.

What It Costs On Older Cards

Be realistic about this before you get excited.

The performance cost was already the main objection on RTX 50 hardware. Published testing of the leaked build showed Control dropping from 71 fps to 35 fps at 4K on an RTX 5070 Ti, and from 91 to 50 on an RTX 5090. That is roughly half your frame rate on the newest cards NVIDIA sells, running the format the model was designed for.

Older cards pay that cost on slower tensor hardware, in a format the model was not built around. Expect the penalty to be worse, and expect it to be worst on RTX 20 series.

This is also a fixed per frame cost rather than something you outrun. The 5090 lost about the same percentage as the 5070 Ti in the same game. A faster card gives you a higher starting frame rate, not a smaller penalty. That logic runs in the other direction too. On an RTX 2060 you are applying a heavy fixed cost to a frame rate that did not have much headroom to begin with.

Why This Matters Beyond The Benchmark Numbers

Two things follow from the requirement collapsing, and the second one is the bigger deal.

The audience got much larger. RTX 20 through 50 covers an enormous share of the PC gaming install base. A story that was interesting to people with current generation flagships is now interesting to people with a five year old card.

It arrived alongside the API barrier falling. In the same window, community work made DLSS 5 reachable in DX9, DX10, DX11 and Vulkan games that never shipped DLSS at all, and then inside emulators. We cover the PCSX2 development and the technique that made it possible separately.

Put together, the practical requirement went from "an RTX 50 card and a game with native DLSS support" to "any RTX card and more or less any game." In under a week. That is a much bigger change than either half sounds on its own.

What This Does Not Change

Some things are exactly as they were.

The image quality argument. Making it run on more hardware does not address the criticism that it reshapes faces and overrides art direction. The Skyrim reaction was not about frame rates.

The legal and safety picture. This still depends on an unreleased NVIDIA library that shipped by accident, and it still uses DLL injection, which is mechanically what cheats do. Our risks article covers why you should keep this away from anything with anti cheat.

Handhelds. Every PC gaming handheld runs AMD or Intel graphics. There is no NVIDIA GPU Windows handheld, so no amount of backwards compatibility on desktop RTX cards reaches the Steam Deck, the ROG Xbox Ally X20, the Legion Go or the Legion C700. Handheld owners are still looking at FSR and XeSS, and cloud streaming remains the only plausible route to DLSS 5 output on a handheld screen.

Our recommendation. Wait. The official release is due Fall 2026 with developer integration, proper masks and optimisation. Everything you would be running today is an unfinished build that escaped by accident.

A Note On Our Earlier Coverage

We said RTX 50 was the whole list. That was accurate to the evidence available on 29 August and it is not accurate now.

The original article carried a note flagging that RTX 40 support was genuinely unknown rather than confirmed absent, which turned out to be the right hedge. We have updated that page rather than leaving it standing, because a stale requirements table is worse than no requirements table.

Frequently Asked Questions

Does DLSS 5 work on RTX 40 series cards?

Yes. RTX 40 was the first generation modders unlocked after the initial RTX 50 only demonstrations, and it works without the numeric format problems that affected older cards. Performance is still heavily reduced compared to running without neural rendering.

Does DLSS 5 work on RTX 30 or RTX 20 cards?

Yes, through an FP16 path. The leaked model targets FP8, which Ampere and Turing tensor cores do not support in hardware, and early attempts on that path performed terribly. Switching to FP16 produced substantially better results on the same cards.

Why did DLSS 5 run so badly on RTX 30 at first?

Because it was being run in FP8, a format those tensor cores have no native support for. RTX 30 Ampere and RTX 20 Turing lack the FP8 and FP4 units that Blackwell has, so the work had to be emulated. Running the model in FP16 instead avoids that.

Can a GTX 1080 or GTX 1660 run DLSS 5?

No. Those cards have no tensor cores, which neural rendering requires. OptiScaler can route some GPUs onto FSR based upscaling and frame generation paths, but that is a different technology and not DLSS 5.

Does DLSS 5 work on AMD or Intel GPUs?

No. DLSS is an NVIDIA technology that runs on NVIDIA tensor cores. Radeon and Arc cards cannot run it at all. AMD's equivalent is FSR and Intel's is XeSS.

Is the performance hit smaller on a faster card?

Not proportionally. It is a fixed per frame cost, so an RTX 5090 lost roughly the same percentage as an RTX 5070 Ti in the same game. A faster card gives you a higher starting point rather than a smaller penalty, which also means older cards have less headroom to absorb it.

Should I try DLSS 5 on my older RTX card?

We would not recommend it. You would be running leaked, unoptimised code in a numeric format it was not designed for, at a heavy frame rate cost, for a visual change many people dislike. The official DLSS 5 release is due in Fall 2026.