The RTX 4090 and RTX 4080 can apparently run NVIDIA’s leaked DLSS 5 Neural Rendering implementation after modders patched its CUDA binaries for Ada Lovelace. But that does not mean either GPU officially supports DLSS 5.
The distinction matters. NVIDIA has not announced DLSS 5 Neural Rendering support for the RTX 40 series, and the current experiments involve an unfinished, leaked implementation rather than a final consumer release.
For RTX 4090 and RTX 4080 owners, the discovery answers one question more clearly than before: Ada Lovelace does not appear to be fundamentally incapable of running DLSS 5. The harder question is whether it can run the technology efficiently enough for NVIDIA to support it.
RTX 4090 and RTX 4080 DLSS 5 support explained
| GPU | Architecture | Unofficial DLSS 5 test | Official support |
|---|---|---|---|
| RTX 4090 | Ada Lovelace | Yes | Not announced |
| RTX 4080 | Ada Lovelace | Yes | Not announced |
| RTX 4080 SUPER | Ada Lovelace | Limited evidence | Not announced |
| RTX 4070 Ti/SUPER | Ada Lovelace | Limited/unknown | Not announced |
| RTX 4070/SUPER | Ada Lovelace | Unknown | Not announced |
| RTX 4060 series | Ada Lovelace | Unknown | Not announced |
| RTX 5090/5080 and other RTX 50 series GPUs | Blackwell | Yes | Designed for it |
So, if the question is “Can an RTX 4090 execute the leaked DLSS 5 Neural Rendering code?”, the answer is apparently yes.
If the question is “Does NVIDIA officially support DLSS 5 on the RTX 4090?”, the answer is no.
Those are two very different claims.
What makes DLSS 5 different from existing DLSS features
DLSS 5 is not simply another version of Super Resolution.
Existing DLSS technologies perform specific jobs. Super Resolution reconstructs a higher-resolution image from a lower-resolution render, Frame Generation creates additional frames, and Ray Reconstruction reconstructs ray-traced lighting information.
Neural Rendering is broader.
NVIDIA’s DLSS 5 concept uses a neural model to transform aspects of the rendered scene, potentially affecting lighting, materials, skin, hair, shadows, reflections, and surface detail. The goal is to shift more visual workload from conventional rendering to AI inference.
That is also why hardware matters more than it did with earlier DLSS versions.
RTX 4090 already has substantial AI hardware
The RTX 4090 is an Ada Lovelace GPU with fourth-generation Tensor Cores, 16,384 CUDA cores, and 24GB of GDDR6X memory. The RTX 4080 also uses Ada Lovelace and fourth-generation Tensor Cores.
In other words, neither GPU lacks dedicated hardware for AI workloads.
The important difference is that NVIDIA positioned DLSS 5 around newer Blackwell hardware. RTX 50 series GPUs use fifth-generation Tensor Cores and introduce newer AI-oriented capabilities that make them better suited to the workload.
That does not automatically make Ada incompatible.
Modders found a way around the apparent compatibility barrier
The breakthrough came from a leaked DLSS Neural Rendering DLL found in an early-access build of NBA 2K27. The file, nvngx_dlssnr.dll, was extracted and examined by modders.
The initial implementation appeared to rely on CUDA binaries containing instructions intended for newer hardware. The RenoDX modding community subsequently patched those binaries so they could execute on Ada Lovelace.
Testing reportedly demonstrated the modified implementation on the RTX 4090 and RTX 4080.
That changes the hardware argument.
Instead of showing that RTX 40 series GPUs simply lack the necessary hardware, the experiment suggests that at least some of the initial barrier involves software, compiler, and instruction compatibility.
But getting the code to execute is only the first hurdle.
The RTX 4090 can run it, but the performance cost is the bigger story
Wccftech reported that an RTX 4090 test saw performance fall from roughly 135–145 FPS to around 83 FPS with the experimental Neural Rendering implementation enabled.
That represents a roughly 39% performance reduction.
For an RTX 4090, that is significant.
The GPU has considerable compute resources and 24GB of VRAM, yet the leaked implementation can still impose a substantial cost. The result should not be treated as a measurement of final DLSS 5 performance because the tested software is unofficial and unfinished, but it illustrates the problem NVIDIA faces with RTX 40 series support.
A feature can be technically executable without being practical.
Neural Rendering has to perform additional inference operations alongside the game’s existing rendering workload. That can consume Tensor Core resources, memory bandwidth, GPU memory, and synchronization time.
Blackwell’s newer AI hardware could therefore make the difference between “it runs” and “it runs efficiently.”
Why RTX 50 series remains the safer bet
Blackwell was built for heavier AI workloads. RTX 50 series GPUs add fifth-generation Tensor Cores, FP4 support, newer RT hardware, and other neural-rendering-oriented capabilities.
That gives NVIDIA considerably more room to optimize an expensive neural-rendering model for the newer architecture.
It also explains why RTX 50 series being the primary DLSS 5 platform does not necessarily prove that RTX 40 series hardware is incapable of the technology.
There is a middle ground:
Ada may be capable of executing DLSS 5, while Blackwell may be capable of executing it at the performance level NVIDIA considers acceptable.
That distinction is probably more important than a simple compatibility list.
What this means for RTX 4090 and RTX 4080 owners
The recent testing makes future RTX 40 series support more plausible, but it does not establish that NVIDIA will provide it.
Several reasons could still keep Neural Rendering limited to RTX 50 series GPUs.
Performance is the obvious one. If an Ada implementation consumes too much GPU performance, NVIDIA may decide the resulting experience isn’t good enough.
Optimization is another factor. The final model could make heavier use of Blackwell-specific AI capabilities.
Power efficiency also matters. Two GPUs may produce the same neural-rendering output while requiring very different amounts of GPU resources.
There is also the less technical issue of product segmentation. NVIDIA has a reason to reserve some of its newest rendering capabilities for its latest architecture.
And official support involves much more than getting a DLL to launch. NVIDIA would need to validate drivers, games, different GPUs, memory configurations, and performance profiles.
RTX 4090 versus RTX 4080
Between the two, the RTX 4090 is the more interesting candidate.
Both cards use Ada Lovelace and fourth-generation Tensor Cores, so the reported success on both supports the idea that the architecture itself is not an absolute blocker.
The RTX 4090, however, has substantially more compute resources and 24GB of VRAM. That gives it more headroom for a computationally expensive workload.
The RTX 4080 can apparently run the patched implementation too, but that doesn’t mean it will deliver the same experience.
The distinction is simple:
RTX 4090: experimentally demonstrated, with more hardware headroom.
RTX 4080: experimentally demonstrated, but with fewer resources available for a demanding neural-rendering workload.
Neither has official DLSS 5 support.
RTX 30 series and RTX 20 series are a different question
The Ada breakthrough should not be extrapolated directly to older RTX generations.
The leaked implementation reportedly contains newer CUDA and FP8-related functionality, while the successful RTX 40 series tests involved modifying binaries specifically to work around incompatible instructions.
That makes an RTX 3090 conclusion premature.
RTX 20 series GPUs are even less certain. They support DLSS and have Tensor Cores, but DLSS 5 Neural Rendering represents a substantially different workload. Leaks currently provide no basis for assuming RTX 20 series cards will support it.
The leaked build is not the final DLSS 5 experience
This point should remain front and center whenever the RTX 4090 tests are discussed.
The current experiments use a leaked implementation extracted from an early-access game build. It has compatibility problems, performance overhead, limited integration, and visual issues.
Some early demonstrations have also shown unusual changes to faces and character appearance.
That means the roughly 39% RTX 4090 performance penalty should not be presented as the expected cost of the final DLSS 5 release.
It is evidence that the current implementation is expensive on Ada, nothing more.
What NVIDIA could do with the RTX 40 series
If NVIDIA decides to support Ada, it doesn’t necessarily have to offer the same Neural Rendering configuration as Blackwell.
A separate model, lower-precision mode, reduced resolution, or reduced neural workload could theoretically let RTX 40 series GPUs offer a different version of the technology.
That would let NVIDIA extend the feature without requiring Ada to match Blackwell’s performance.
But this remains speculation until NVIDIA provides an official compatibility announcement.
For now, the evidence supports a narrower conclusion.
The RTX 4090 and RTX 4080 can apparently be made to run leaked DLSS 5 Neural Rendering code.
NVIDIA has not confirmed official RTX 40 series support.
The current Ada implementation carries a substantial performance cost.
RTX 50 series remains the architecture DLSS 5 is designed around.
That puts the RTX 4090 in an interesting position. It is not an officially supported DLSS 5 GPU, but it also cannot simply be dismissed as hardware incapable of running the technology.
The next meaningful development will not be another modded benchmark. It will be NVIDIA’s final compatibility list — and whether the company considers RTX 40 series performance good enough to make Neural Rendering a supported feature rather than an RTX 50 series-only feature.
