A leaked build of NVIDIA’s next-generation upscaler is running on hardware it was never meant to touch. Reports from TechPowerUp and GameGPU describe a patched DLSS 5 DLL, pulled from an early build of NBA 2K27, now working on GeForce RTX 4000 “Ada Lovelace” cards, despite NVIDIA positioning DLSS 5 as a GeForce RTX 5000 “Blackwell” exclusive. The fix did not come from NVIDIA. It came from a single modder with a hex editor and a theory about why the DLL was locked down in the first place.
What happened: a leaked DLL meets a modder’s patch
The story starts with a leak, not a release. An early access build of NBA 2K27 shipped with a DLSS 5 component still bundled inside it, a mistake that has become almost routine in an industry where unreleased NVIDIA features keep surfacing in unfinished game builds months ahead of schedule. The file at the center of the story is nvngx_dlssnr.dll, NVIDIA’s neural rendering library, and outlets covering the leak describe it as an early implementation of DLSS 5 rather than a finished, shipping version.
On its own, that leak would have been a curiosity. What turned it into news is that DLSS 5 was clearly built to run only on GeForce RTX 5000 “Blackwell” GPUs, and a modder going by Uncle Burrito found a way around that restriction. Within days, TechPowerUp and GameGPU were both reporting that the patched DLL ran on GeForce RTX 4090 and RTX 4080 cards, two GPUs built on the older Ada Lovelace architecture that NVIDIA never listed as DLSS 5 hardware.
NVIDIA has not commented on the mod, and nothing about this patch turns DLSS 5 into an officially supported feature on RTX 4000 cards. It is an unofficial workaround running on leaked, pre-release code, and it could stop working the moment NVIDIA ships a real driver or a signed, finished build of DLSS 5.
Inside the DLL: sm_120 Blackwell code meets sm_89 Ada silicon
To understand why this patch was even possible, it helps to know what GPU code actually looks like once it is compiled. NVIDIA compiles CUDA binaries for a specific compute capability, essentially a version number tied to a GPU architecture, and a binary built for one compute capability will not simply run on a different one without either a recompile or a fallback path baked in.
The leaked DLSS 5 DLL contained CUDA binaries compiled for compute capability sm_120, the identifier tied to Blackwell, the architecture behind the RTX 5000 series. According to the reporting, the DLL contained no binaries for sm_89 (Ada Lovelace, the architecture behind the RTX 4000 series) and none for the older Ampere generation either. In plain terms, the file was built exclusively for Blackwell and nothing else, which is exactly what you would expect if NVIDIA genuinely intends DLSS 5 as a Blackwell-generation feature rather than a marketing restriction layered on top of hardware that could already run it.
Why Uncle Burrito’s approach worked
Uncle Burrito, a modder previously known for work in the RTX Remix community, described the process in terms that make it sound almost mechanical rather than exotic. The approach, as reported, was to look at which binaries the DLL called on, identify the ones that had no Ada-compatible counterpart, and swap or patch them so the DLL would call binaries the RTX 4000 architecture could actually execute. That is a binary-compatibility patch, not a rewrite of NVIDIA’s neural network, and it is the same category of trick modders have used for years to unlock frame generation and ray reconstruction features on GPUs NVIDIA didn’t originally list as supported.
| Item | Blackwell (RTX 5000 series) | Ada Lovelace (RTX 4000 series) |
|---|---|---|
| Compute capability in leaked DLL | sm_120 (present, official target) | Not present in original leak |
| NVIDIA’s stated DLSS 5 support | Yes, officially positioned as Blackwell-only | No official support |
| FP8 tensor core support | Yes, native | Yes, native |
| Status after Uncle Burrito’s patch | Runs as originally leaked | Runs unofficially on RTX 4090 and RTX 4080 |
| Older Ampere (RTX 3000) support | Not applicable | Not present in leak or patch |
Why FP8 precision is the real reason this worked
The detail that made the patch technically possible, rather than just theoretically possible, is precision format. Reporting on the leak describes the DLSS 5 neural network running in FP8, an 8-bit floating point format that trades numerical range for speed and memory efficiency. FP8 has become a standard tool in modern AI inference because it lets a neural network run with a fraction of the memory footprint and compute cost of FP16 or FP32, at the cost of some precision that well-trained models can tolerate.
Ada Lovelace’s tensor cores support FP8 natively, in hardware, not through software emulation. That is the crux of the entire story: DLSS 5’s neural model doesn’t need silicon that only exists in Blackwell. It needs FP8 tensor core throughput, and RTX 4000 already has that. NVIDIA restricting DLSS 5 to Blackwell on paper doesn’t change what the Ada Lovelace tensor cores are physically capable of running once someone points compatible binaries at them.
That does not mean RTX 4000 and RTX 5000 would perform identically running DLSS 5. Blackwell has newer tensor core designs, more of them, and higher throughput per core, so even with FP8 support in common, an RTX 5090 running DLSS 5 natively would be expected to have more headroom than an RTX 4090 running a patched build. What the FP8 overlap proves is narrower but still significant: the restriction to Blackwell is a business decision layered on top of hardware capability, not a hard technical wall.
The demo: DLSS 5 running Control on an RTX 4090
Coverage of the mod includes footage of the patched DLL running in Control, a game long used as an NVIDIA showcase title for ray tracing and upscaling technology. According to the reporting, the demo shows the patched DLSS 5 build running at approximately 82 frames per second on a GeForce RTX 4090. That single data point confirms the DLL executes and renders frames on Ada Lovelace hardware, which is itself the headline: a pre-release, Blackwell-targeted DLSS build producing a playable frame rate on a two-generation-old architecture.
What isn’t confirmed is a clean, apples-to-apples performance comparison against DLSS 4 or native rendering on the same hardware. Leaked, pre-release AI models are rarely optimized, and a patched binary path is not the same as a build NVIDIA has tuned and validated for a specific GPU. Readers should treat the frame rate as evidence the mod works, not as a preview of what DLSS 5 will look like on RTX 4000 once, or if, NVIDIA ever ships it there officially.
NVIDIA’s official position: Blackwell only, for now
Nothing about this leak changes NVIDIA’s public stance. DLSS 5 has been positioned since it first surfaced in NVIDIA’s own materials as a feature tied to the RTX 5000 series, in the same way DLSS 4’s Multi Frame Generation launched as a Blackwell-only capability. NVIDIA has not issued a statement responding to the Uncle Burrito patch specifically, and the company has historically declined to comment on unofficial mods that unlock features outside their stated hardware requirements.
That silence is itself informative. NVIDIA typically does not rush to shut down community workarounds unless they threaten revenue directly, such as mining-limiter bypasses once did. A DLL patch that requires digging leaked pre-release files out of a sports game, replacing internal binaries by hand, and running on unsupported, unvalidated code is not a mass-market threat to RTX 5000 sales. It is a proof of concept for enthusiasts, and NVIDIA can afford to watch it play out before deciding whether to respond.
A familiar pattern: NVIDIA locking software to new silicon
This is not the first time NVIDIA has drawn a hardware line around a DLSS feature that the previous generation could arguably still run. Frame generation launched as an RTX 40-series exclusive when DLSS 3 debuted, a decision that drew criticism at the time from RTX 30-series owners who argued their tensor cores could handle the workload, even if not as efficiently. Multi Frame Generation followed the same script with DLSS 4 and Blackwell. Each time, NVIDIA’s reasoning centered on the newer architecture’s Optical Flow Accelerator and tensor core improvements delivering a meaningfully better result, not just a functional one.
DLSS 5 restricted to Blackwell fits that same pattern, and the FP8 overlap with Ada Lovelace only sharpens the comparison. It suggests NVIDIA’s generational locks are, at least in part, product segmentation dressed up as a hardware requirement, a strategy that gives buyers a reason to upgrade even when the older card is technically capable of a lesser version of the same feature.
| DLSS generation | Headline feature | NVIDIA’s official hardware requirement |
|---|---|---|
| DLSS 3 | Frame Generation | RTX 40-series (Ada Lovelace) and newer |
| DLSS 4 | Multi Frame Generation, transformer model upscaling | RTX 50-series (Blackwell), with transformer upscaling later extended to older RTX cards |
| DLSS 5 | Neural rendering (per leaked build) | RTX 50-series (Blackwell) only, per current leak |
| DLSS 5, patched | Same neural rendering model | Confirmed running unofficially on RTX 4090 and RTX 4080 |
How the patch works, in plain terms
Nobody has published the patched DLL’s source, and shattered.io is not distributing or reverse-engineering NVIDIA’s code here. But the reported method, as described across outlets covering the mod, breaks down into a straightforward sequence rather than anything requiring a fresh AI model.
1. Extract nvngx_dlssnr.dll from the leaked NBA 2K27 build
2. Identify the CUDA binaries (cubins) embedded in the DLL
3. Flag binaries compiled only for sm_120 (Blackwell)
that have no Ada-compatible (sm_89) equivalent
4. Replace or patch those binaries so the DLL's
Ada Lovelace code path is reachable
5. Load the patched DLL in a compatible game/renderer
on an RTX 4000-series GPU
6. Confirm execution and frame output (Control, RTX 4090)The neural network itself, the actual DLSS 5 model, was not retrained or rebuilt for Ada Lovelace. It is the same model NVIDIA compiled for Blackwell, executing on Ada Lovelace tensor cores because both architectures share native FP8 support. That is a meaningfully smaller task than building a new upscaler from scratch, which is part of why one modder was able to get results in days rather than months.
Market impact: what this means for RTX 4000 owners right now
For the millions of gamers still running RTX 4090, RTX 4080, and other Ada Lovelace cards, the practical takeaway is limited but real. This is not a driver update, not a supported feature, and not something NVIDIA will patch support tickets around if it breaks a game. It is a leaked, pre-release file paired with a hand-patched binary, running in one demonstrated title. Anyone installing it is running unsigned, unofficial code from a modder, on top of an NBA 2K27 file NVIDIA never intended to ship publicly in that state.
The bigger impact is on expectations. Every time a locked NVIDIA feature turns out to run on older hardware once someone patches the binary, it adds pressure on NVIDIA to eventually walk the restriction back, the way DLSS 4’s transformer-based upscaling model was later extended to RTX 20, 30, and 40-series cards after initially shipping tied to newer hardware messaging. If that pattern repeats, RTX 4000 owners have a real, if unofficial, precedent for expecting at least some scaled-down version of DLSS 5 to arrive on their cards eventually, even if Blackwell keeps the full feature set.
For NVIDIA, the sales calculus is more complicated than a simple “this hurts Blackwell sales” read. The mod is not turnkey, it depends on a leak NVIDIA didn’t intend to happen, and it may stop working entirely once the leaked pre-release code is superseded by a finished, signed DLSS 5 release. That gives NVIDIA room to let the story play out rather than treat it as an urgent threat to RTX 5000 revenue.
Competitive landscape: how AMD and Intel fit into this
NVIDIA’s generational lock strategy stands out more sharply against how its two main rivals in upscaling have approached hardware requirements. AMD’s FSR has generally prioritized broad hardware compatibility, including running on non-AMD GPUs and older cards, as a competitive differentiator against DLSS’s NVIDIA-only, and often generation-specific, requirements. Intel’s XeSS has followed a similar path, with a cross-vendor fallback path alongside a faster, Intel-optimized path on its own Arc GPUs.
That contrast is exactly why a story like this one resonates. Every time NVIDIA locks a headline feature to its newest cards and a modder proves the older generation can run a version of it anyway, it hands AMD and Intel a talking point about openness versus planned upgrade cycles. Neither AMD nor Intel has publicly commented on the DLSS 5 leak, and there is no indication either company is accelerating a competing feature in direct response, but the optics of the leak land at a moment when both rivals are actively marketing broader hardware support as a selling point against NVIDIA’s ecosystem.
The risk side: why this isn’t something to install casually
It’s worth being direct about the downside here. This patch runs code extracted from a leaked, pre-release game build, modified by a third party, executing a neural rendering pipeline NVIDIA has not validated for Ada Lovelace hardware. That combination carries real risk: instability, crashes, incompatibility with future driver updates, and no support path if something goes wrong, since NVIDIA support will not troubleshoot a modded DLL running unofficial code.
There is also a legal and distribution gray area worth naming plainly. The underlying DLL is NVIDIA’s intellectual property, extracted from a game that was never meant to expose it publicly. Modding communities have operated in this space for years without major legal action from NVIDIA, but that history is not a guarantee, and anyone downloading a patched file from a third-party source is trusting that source not to have bundled anything malicious alongside a legitimately interesting technical trick.
Historical context: leaks have driven DLSS coverage before
NVIDIA’s DLSS features have leaked ahead of official announcements more than once, usually surfacing inside game files, driver packages, or SDK updates before NVIDIA formally confirms them. That pattern has effectively turned data-mining leaked game builds into a recurring source of NVIDIA news, one that outlets like TechPowerUp, VideoCardz, and Wccftech regularly cover well before NVIDIA’s own announcement calendar catches up. The NBA 2K27 leak fits squarely into that history: a AAA game shipped a debug or early-access build with a component that wasn’t supposed to be publicly accessible yet, and the enthusiast press found it within days.
What makes this particular leak different is that it produced a working, demonstrable mod rather than just a naming reveal or a spec sheet. Most DLSS leaks tell you a feature exists before NVIDIA says so. This one let someone actually run it on hardware NVIDIA says doesn’t support it, which is a meaningfully bigger story for anyone trying to gauge how real the Blackwell-only restriction actually is.
Predictions: where this goes from here
- NVIDIA stays quiet in the near term. Expect no official statement addressing the Uncle Burrito patch directly, consistent with NVIDIA’s pattern of not engaging publicly with unofficial mods that don’t threaten immediate sales.
- The leaked build gets patched out. NBA 2K27’s shipping release is likely to replace or lock down the exposed DLSS 5 DLL before or shortly after launch, cutting off the easiest path to the current mod.
- A scaled-down DLSS 5 path eventually reaches RTX 4000. Following the precedent set when DLSS 4’s transformer upscaling model was extended to older RTX cards after launch, expect pressure to grow for NVIDIA to officially support at least some DLSS 5 functionality on Ada Lovelace, even if Blackwell keeps exclusive features.
- More modders will chase the same trick on other titles. Once one leaked DLL has been shown to be patchable, expect the modding community to look for the same DLSS 5 component in other early-access or beta builds circulating ahead of a wider release.
- AMD and Intel will lean into the contrast rhetorically. Expect continued marketing emphasis from both companies on broad hardware compatibility for FSR and XeSS, using stories like this one as an implicit contrast point even without naming NVIDIA directly.
What to watch next
The next concrete signal will come from NBA 2K27’s actual launch build. If the shipping version of the game no longer contains an exploitable DLSS 5 DLL, that confirms 2K and NVIDIA treated the exposure as a mistake to be closed rather than an early preview to leave in place. The second signal to watch is whether NVIDIA’s next driver release contains any code that specifically blocks or breaks the patched execution path, which would be the clearest sign NVIDIA considers the mod worth actively countering rather than ignoring.
Longer term, the real test is whether NVIDIA ever ships an official, supported version of any DLSS 5 functionality for RTX 4000 hardware. Given that the company has walked back generational exclusivity before, most notably with DLSS 4’s transformer model reaching older RTX cards after launching tied to newer messaging, RTX 4000 owners have a reasonable basis for that expectation. Whether it happens on NVIDIA’s terms, or gets forced by more mods like this one, is the open question.
Frequently asked questions
Is DLSS 5 officially supported on RTX 4090 or RTX 4080?
No. NVIDIA has positioned DLSS 5 as a GeForce RTX 5000 “Blackwell” feature. The version running on RTX 4000 cards is an unofficial patch applied to a leaked, pre-release DLL, not a supported NVIDIA release.
Where did the leaked DLSS 5 file come from?
Reports trace the leaked nvngx_dlssnr.dll file to an early access build of NBA 2K27, which shipped with the DLSS 5 component still bundled inside it.
Who is Uncle Burrito?
Uncle Burrito is a modder previously associated with the RTX Remix modding community, credited with identifying the Blackwell-only CUDA binaries inside the leaked DLL and replacing them with Ada Lovelace-compatible code so the file would run on RTX 4000 GPUs.
Why can Ada Lovelace GPUs run a Blackwell-targeted AI model at all?
The DLSS 5 neural network runs in FP8 precision, a format that Ada Lovelace tensor cores support natively in hardware. Because both Blackwell and Ada Lovelace can execute FP8 workloads, the same trained model can run on either architecture once compatible binaries are in place, even though NVIDIA has not enabled that path officially.
Is it safe to install the patched DLSS 5 DLL?
It carries real risk. The file originates from a leaked, pre-release game build, has been modified by a third party, and is not validated or supported by NVIDIA. There is no official support path if it causes instability, and downloading modified DLLs from unofficial sources always carries some risk regardless of the underlying technical trick.
Does this mean RTX 3000 “Ampere” GPUs can run DLSS 5 too?
Not based on current reporting. The leaked DLL did not contain binaries for Ampere, and Ampere’s tensor cores lack the same level of native FP8 support found in Ada Lovelace, making a similar patch significantly harder to achieve, if it’s possible at all.
Will NVIDIA block this patch in a future driver?
NVIDIA has not commented on the mod. It’s plausible a future driver or the finished DLSS 5 release changes the DLL in a way that breaks the current patch, especially once NBA 2K27’s leaked build is superseded by its official launch version.
How does this compare to AMD’s FSR and Intel’s XeSS hardware requirements?
AMD’s FSR and Intel’s XeSS have generally emphasized broader hardware compatibility, including support across older GPUs and, in FSR’s case, non-AMD hardware, in contrast to NVIDIA’s pattern of tying headline DLSS features to its newest GPU generation at launch.
