A leak turned into a rapid compatibility experiment

A leaked build of NVIDIA’s forthcoming DLSS 5 neural-rendering technology has become an unusually public test of how far graphics features can be pushed beyond their intended hardware. The library reportedly surfaced in an early-access build of NBA 2K27 and was soon adapted by members of the RenoDX modding community for games outside its original context.

The latest step was to make it initialise on GeForce RTX 30-series GPUs, NVIDIA’s Ampere generation. That achievement should be treated narrowly: it demonstrates that the software can be made to execute, not that it delivers a viable feature for players. The available results on August 30, 2026 point to severe losses in frame rate, much higher render latency and highly inconsistent behaviour between games.

This matters because DLSS has traditionally been understood as a way to improve the performance-to-image-quality trade-off. The current tests instead show an experimental neural-rendering workload that can consume most of an Ampere GPU’s available performance. For owners of RTX 30-series cards, the practical conclusion is straightforward: the leaked implementation is an interesting proof of concept, not an upgrade path.

Early numbers show the scale of the problem

Reported community tests vary in methods, graphics settings, resolutions and game integrations, so they are not a substitute for a controlled benchmark suite. Even so, the direction of the results is consistent.

An RTX 3070 Laptop GPU test in Kingdom Come: Deliverance II reportedly saw render latency rise from roughly 29 milliseconds to more than 3.3 seconds after neural rendering was enabled. Separately, an RTX 3080 running Deep Rock Galactic was reported to fall from about 130 frames per second to four frames per second. Other claims placed RTX 3050 and RTX 3060 Ti results at around one frame per second in games including Control, FiveM and Kingdom Come: Deliverance II.

There are exceptions, but they reinforce rather than overturn the main point. Some high-end Ampere cards reportedly reached roughly 20 to 40 frames per second in particular games. Those outcomes often involved substantial compromises, such as low internal resolution, aggressive upscaling modes, lowered game settings or visible noise and shimmering. A result that is technically playable after reducing the source image heavily is not equivalent to broad support at normal quality settings.

Frame rate alone also understates the issue. A visually sophisticated processing stage is of limited use in an interactive game if it adds enough delay or uneven frame pacing to make input feel disconnected. The latency reports are therefore at least as significant as the headline frame-rate collapses.

Why Ampere is an especially difficult target

The apparent bottleneck is closely tied to hardware capabilities. GeForce RTX 30-series products use Ampere GPUs with third-generation Tensor Cores. NVIDIA’s technical documentation lists their matrix formats including FP16, BF16, TF32 and integer modes, but not native FP8 floating-point Tensor operations.

FP8 acceleration arrived in NVIDIA’s subsequent Ada Lovelace generation, used by the RTX 40 series. NVIDIA described Ada’s fourth-generation Tensor Cores as containing an FP8 Tensor Engine. That distinction helps explain why modders first made the leaked library work on RTX 40-series hardware and why Ampere support has been much more punishing.

The community’s Ampere work reportedly relies on an experimental FP16-oriented route rather than the model’s preferred path. Inference can sometimes be translated, emulated or otherwise adapted to a different numerical format, but those workarounds can eliminate the speed advantage that made the model plausible in real time. The outcome is a workload that runs but no longer fits comfortably inside a game’s frame-time budget.

That is an important distinction in the debate around generational feature support. A vendor restriction is not automatically proof that older hardware is incapable of executing a function. Equally, bypassing a check does not prove that the feature can be supported responsibly. Stable drivers, predictable image quality, manageable latency and acceptable performance are all part of meaningful support.

What the experiments do — and do not — establish

The leaked files and patches provide limited evidence about NVIDIA’s eventual product plans. They do suggest that the technology is computationally demanding and that its initial implementation includes code paths targeted at newer architectures. They do not establish final image quality, finished performance, game-by-game compatibility or the GPU support policy for the commercial release.

The library being tested is neither a final public SDK nor an official game integration. It has been inserted into titles through community tooling, and it may lack the optimisations, safeguards and per-game tuning that a released implementation would receive. Results can also be affected by the particular model settings, the chosen render resolution, CPU performance, driver version and the quality of the mod’s integration.

The comparison with RTX 40-series experiments is useful but must also be cautious. Ada’s FP8 capability gives it a more suitable execution path, yet that does not mean every RTX 40-series card will necessarily receive official access. NVIDIA has not publicly confirmed a full DLSS 5 compatibility matrix. Until it does, claims about future support for Ada or Ampere remain speculation.

A clearer message for RTX 30 owners

The experiment has value for developers and enthusiasts because it exposes the relationship between AI models, numerical precision and GPU architecture. It also highlights how rapidly a leaked component can be adapted by a technically capable community.

For consumers, however, the message is less dramatic. RTX 30-series hardware remains capable of established DLSS features and its Tensor Cores remain useful for many AI-assisted graphics tasks. The current neural-rendering patch does not change that. It instead illustrates that future AI graphics features may depend as much on specialised data formats and inference throughput as on conventional shader performance.

Owners should not make purchasing or upgrade decisions based on isolated screenshots, social-media clips or unrepeatable tests from a leaked build. The meaningful comparison will come only when NVIDIA releases a supported version, documents its hardware requirements and independent reviewers can measure image quality, latency and performance using consistent settings. Until then, Ampere’s apparent ability to run the leaked DLSS 5 code is best regarded as a technical curiosity with an emphatic performance warning.

Sources