Several reported results are dramatic in the wrong direction: frame rates dropping into single digits and render latency climbing from milliseconds to multiple seconds. A small number of higher results exist, but their settings matter greatly. One reported 41 FPS result in Cyberpunk 2077, for example, used 720p resolution and the Ultra Performance upscaling mode. That is an intriguing proof that the patched code can sometimes run, not evidence of broadly playable performance.
This is not the DLSS 5 launch
NVIDIA publicly unveiled DLSS 5 on March 16, 2026, and said the technology would arrive in fall 2026. The current reports do not concern that future consumer release. They concern a pre-release neural-rendering DLL that appeared in an early-access build of NBA 2K27 and was then modified by third parties.
That provenance changes how every claim around the experiment should be read. A pre-release DLL may be incomplete, tied to a particular game integration, compiled for a particular GPU target, or simply not representative of the code NVIDIA ultimately ships. There is no indication in the available record that NVIDIA authenticated the leaked component, endorsed its use, or described it as final.
It also means that three separate questions must not be blended together:
- Can a patched DLL be made to load on a GPU?
- Can it produce an image and run in a particular game setup?
- Will NVIDIA officially support that GPU in the released DLSS 5 product?
The recent reports provide limited evidence for the first two questions on some older RTX cards. They answer neither the third question nor the more important everyday question: whether players can expect a smooth, dependable experience.
As of August 31, NVIDIA had not publicly set out which GeForce generations would receive official DLSS 5 support. Claims that the technology is officially limited to RTX 50-series hardware therefore go beyond the available evidence. Equally, the fact that modders can force a build to run on older hardware should not be mistaken for a promise of official compatibility.
What the Ampere results actually show
The RTX 30-series results reported so far are best understood as isolated field observations, not a conventional benchmark suite. They come from different games, different GPUs, and unclear combinations of resolution, presets, drivers, CPU hardware, game builds, and patch versions. That makes them valuable warning signs, but poor material for predicting an outcome on a particular PC.
The warnings are nevertheless severe. One reported RTX 3070 Laptop test saw render latency rise from roughly 29 ms to 3,326 ms. Put plainly, the latter is more than three seconds of render latency, a result far outside anything suitable for interactive play. In another report, an RTX 3080 running Deep Rock Galactic reportedly fell from around 130 FPS to 4 FPS with the experimental setup.
Those examples establish that patched pre-release DLSS 5 code can perform disastrously on Ampere. They do not establish that every RTX 30-series card, every game, or every eventual DLSS 5 implementation will behave identically. Terms such as “most games” imply a controlled and sufficiently broad test sample that is not available here.
There are counterexamples. Reports include results above single digits, with Tom’s Hardware describing a figure of up to 41 FPS in Cyberpunk 2077 on an RTX 30-series card. But the reported configuration was 720p with Ultra Performance upscaling—conditions that sharply limit what can be inferred about image quality, ordinary monitor resolutions, or the performance users might see with less aggressive settings.
This is why a single FPS number is not enough. When an upscaler is set to Ultra Performance at 720p, the internal render workload can be very different from a 1080p, 1440p, or 4K configuration using a higher-quality preset. A result that merely demonstrates execution can still be unsuitable for the way most people play games.
Why FP8 is a plausible explanation, but not a verdict
A leading theory behind the Ampere struggles concerns tensor-core precision support. NVIDIA’s GA10x Ampere documentation lists FP16, INT8, INT4, binary, TF32, and BF16 tensor formats, but not FP8. NVIDIA describes Ada Lovelace tensor cores, used in the RTX 40 series, as supporting FP8 precision.
That hardware difference gives the FP8 explanation real technical weight. If the pre-release model or its kernels expect efficient native FP8 execution, an Ampere card may have to use an inefficient alternative path, emulate work through other operations, or encounter code that was never meaningfully tuned for its architecture. Any of those situations could help explain extreme latency and frame-rate losses.
But “plausible” is not the same as “proven decisive cause.” The available evidence does not isolate FP8 as the sole reason for Ampere’s performance. In fact, the RTX 40-series experience shows why a simpler explanation is inadequate.
Ada hardware has FP8 support, yet the leaked library still required patches before it could run on an RTX 4080. Tom’s Hardware reported independently verifying such a patch. The stated obstacle was incompatible CUDA instructions associated with Blackwell-targeted binaries. That demonstrates another layer of the problem: binary and architecture compatibility can prevent execution even where a GPU supports a relevant numerical format.
The poor Ampere results could therefore reflect a combination of factors:
- The absence of native FP8 support on Ampere.
- Blackwell-specific CUDA binaries or other architecture assumptions.
- Immature or incomplete pre-release code.
- The quality and overhead of the unofficial patches.
- Game-specific integration differences.
- Unreported system and graphics settings.
Without controlled comparisons that change one of those variables at a time, assigning a precise share of blame to FP8—or to any other one factor—would be speculation.
RTX 40 compatibility changes the narrative, not the support policy
The verified RTX 4080 patch is an important piece of the story. It indicates that the leaked library was not inherently confined to the newest hardware once modders worked around incompatible instructions. It also undermines the idea that the current experiment can be reduced to a clean “new-generation-only” divide.
Yet it does not establish official RTX 40-series support for DLSS 5. A patched build running in an experimental setting is a demonstration of technical possibility, not a product commitment. Official support normally entails validated drivers, tested game integrations, predictable quality behavior, compatibility across supported configurations, and a vendor willing to stand behind the result. None of that follows automatically from a community patch.
For RTX 30 owners, the same principle is even more important. The experiment demonstrates that at least some barriers can be bypassed. Its performance reports suggest that bypassing those barriers does not make the resulting experience viable.
What Windows gamers should do with this information
The immediate practical takeaway is restraint. RTX 30-series owners should not treat the leaked-DLL reports as a reason to expect a no-cost, near-term DLSS 5 upgrade. Nor should they use the worst reported outcomes as conclusive evidence that an eventual official release cannot work on Ampere. Both claims run ahead of the data.
A sensible approach is to separate today’s test from future buying or upgrade decisions. If an Ampere system is performing well in current games with its existing upscaling and rendering options, the reports provide no evidence that replacing a working setup solely in anticipation of DLSS 5 is necessary. Conversely, users evaluating a new GPU should avoid assuming future DLSS 5 eligibility on any generation until NVIDIA publishes a supported-GPU policy.
For Windows users who regularly tune game settings, the reported figures are also a reminder that neural-rendering labels do not guarantee extra speed. A technology intended to improve visual output or performance can become a major performance penalty when the model, binary, driver, GPU architecture, and game integration are not aligned. The right question is not whether a DLL can be loaded, but whether the complete pipeline produces stable frame times, usable image quality, and reliable behavior in a real game.
Until there are repeatable tests with disclosed settings—and, more importantly, NVIDIA’s shipping code and official support details—the safest conclusion is narrow. Modders have shown experimental DLSS 5 compatibility on RTX 30-series hardware, but the current performance evidence ranges from heavily compromised to unusable. It is an early compatibility experiment, not a preview of the final DLSS 5 experience on Ampere.
The information still needed
A meaningful assessment will require more than screenshots or isolated FPS reports. Useful testing would disclose the exact GPU, laptop or desktop configuration, driver version, CPU, game build, display resolution, internal render resolution, quality preset, ray-tracing settings, patch version, and repeatable test route. It would compare the patched library against the game’s existing rendering path rather than presenting one number without a baseline.
Even then, tests of the leaked DLL would answer only questions about that DLL. The decisive evidence will come after the planned fall release, when NVIDIA can identify supported architectures and reviewers can test the production implementation across games and settings.
For now, the story is not that DLSS 5 has arrived on older GeForce cards. It is that a pre-release component has been made to run beyond its apparent target environment, exposing substantial compatibility and performance problems along the way. That is useful technical evidence—but it is not yet a feature Windows gamers can rely on.