DLSS 5 on AMD Now Runs Across Three Radeon Generations, but the Hardware Gap Remains

Daniel Blanco has expanded DLSS 5 on AMD from one experimental effect into a four-part graphics suite across three Radeon generations. The October 17 Alpha 0.8.0 release adds DLSS Frame Generation to an existing collection that includes Super Resolution, Ray Reconstruction, and Neural Rendering.

That is a significant reversal for features marketed around Nvidia’s RTX hardware. Blanco’s DLSS-NR-on-AMD project now targets Radeon RX 9000, RX 7000, and RX 6000 cards, although support differs by architecture. It also turns Nvidia’s separate technologies into selectable modules inside a single manager.

The technical achievement does not erase the hardware divide. The project remains unofficial, several results come from developer testing, and early independent trials have exposed crashes or severe performance losses. Nvidia still controls the original models, supported hardware, and game integrations.

What changed is access. Radeon owners can now experiment with nearly every major DLSS component through one framework, even when AMD’s official alternatives have narrower hardware or game support. The argument has shifted from whether these features can run on Radeon to whether they run well enough to matter.

DLSS 5 on AMD Becomes a Full Graphics Suite

Alpha 0.8.0 changes DLSS-NR-on-AMD from a Neural Rendering experiment into a modular package covering Nvidia’s main DLSS features.

Blanco first released the project on September 3, the same day Nvidia made DLSS 5 available in NBA 2K27. Its original purpose was narrower: run Nvidia’s new Neural Rendering model on AMD hardware through a Radeon-specific neural network runtime.

Neural Rendering is the unusual component. Nvidia describes it as a final rendering stage that uses a generative model to enhance lighting, skin, materials, reflections, and other visual details. Unlike ordinary upscaling, it changes the final appearance rather than only reconstructing a higher-resolution frame.

Blanco’s October update adds the surrounding technologies needed to build a broader DLSS pipeline. According to the project’s version 0.8.0 notes, the package now includes DLSS Super Resolution 4.5, DLSS Ray Reconstruction, DLSS Frame Generation, and Neural Rendering.

Super Resolution reconstructs a higher-resolution image from a lower-resolution input. Ray Reconstruction replaces several hand-tuned denoisers with a learned model that rebuilds ray-traced detail. Frame Generation creates additional frames between conventionally rendered ones.

Version 0.8.0 supports up to 4X Multi Frame Generation when a game provides the required DLSS Frame Generation inputs. The designation means the pipeline can display as many as four frames for each conventionally rendered frame, including generated output.

The update also accepts Nvidia’s 310.9.16 Neural Rendering files and bundles FSR 4.1.1 upscaling. Bundling one FSR version gives the mod a consistent foundation instead of depending on the particular implementation shipped with each game or driver.

Blanco groups these features inside what he calls a Neural Rendering Supermod. Users can choose modules through an overlay rather than installing a separate workaround for every feature. The project’s NeuralHub manager scans for games, installs the loader, centralizes large model files, and keeps installed components updated.

The manager matters because the earlier installation process demanded more manual file handling. Users had to place libraries into individual game directories and troubleshoot each title separately. NeuralHub reduces that friction, although it does not turn the project into a supported driver feature.

A compatible game must use DirectX 12 and expose FSR, DLSS, or Intel XeSS inputs. After installing the loader, a player enables one of those technologies in the game. The mod intercepts that pipeline and routes the available inputs through the selected output modules.

Users must supply their own Nvidia DLL files for the relevant DLSS functions. FSR comes bundled with the manager, but the Nvidia libraries do not. That requirement keeps the project dependent on files distributed with games supporting those technologies.

The architecture support also needs qualification. The project says every main DLSS feature works across RDNA 4 and RDNA 3, which power Radeon RX 9000 and RX 7000 cards. RDNA 2 support remains under testing, with Neural Rendering available before several other modules reach that generation.

That makes “three generations” accurate at the project level, but not equivalent to complete feature parity on every supported card. Radeon RX 6000 owners should treat current compatibility as partial and evolving.

The release adds a native DirectX 12 mode intended to reduce stalls associated with interoperability between graphics and compute systems. It also brings DirectX 12 execution to several features on RDNA 3, alongside faster loading and general network optimizations.

These are not merely menu additions. Blanco says the runtime is a ground-up reimplementation of the neural network execution path for AMD architectures, including HIP kernels, memory layouts, and game integration. The project states that it does not translate Nvidia code during runtime.

That distinction explains why the update represents more than a brand-name toggle. The system must take models designed around Nvidia’s stack and execute their workloads through Radeon-compatible code while preserving the inputs expected by each game.

The new package therefore creates the article’s central tension. Software has made Nvidia’s suite portable enough to run elsewhere, but successful execution is not the same as native efficiency, reliability, or official support.

Radeon Access Puts Pressure on AMD’s Feature Strategy

The mod pressures AMD because it exposes a demand for broader compatibility even when Radeon already has alternatives to most DLSS functions.

AMD is not starting from an empty graphics stack. Its FSR family covers upscaling and frame generation, while FSR “Redstone” adds machine-learning-based Ray Regeneration and Radiance Caching. Radeon Anti-Lag 2 addresses latency in supported games.

The gap is uneven availability. AMD says its current FSR feature suite depends on supported games, Radeon generations, and specific integration requirements. Not every FSR feature reaches every Radeon owner through the same path.

FSR Redstone’s full collection targets Radeon RX 9000 hardware. AMD lists machine-learning upscaling for RX 7000 and RX 9000 cards, with RX 6000 support scheduled for 2027. Older FSR versions cover more hardware but do not offer the same learned models.

Blanco’s project attacks that fragmentation from another direction. It uses existing FSR, DLSS, or XeSS integration points as entryways, then lets players select from several reconstructed Nvidia and AMD modules.

This does not mean the mod supports every DirectX 12 game. A title still needs a suitable upscaling interface and the required inputs for the chosen feature. Ray Reconstruction needs ray-tracing data, while DLSS Frame Generation needs a compatible Frame Generation path and currently requires GPU spoofing.

Still, the approach widens the practical search space. Instead of waiting for a developer to add one specific vendor feature, modders can attempt to reuse information already available in a game’s rendering pipeline.

That matters because PC gamers increasingly treat image reconstruction and generated frames as part of a graphics card’s value. Raw rasterization performance no longer tells the entire story when demanding ray tracing can depend on upscaling, denoising, and frame synthesis.

Nvidia has strengthened that association by packaging its technologies under DLSS. A game carrying DLSS support can expose Super Resolution, Ray Reconstruction, and Frame Generation through related interfaces, although hardware eligibility varies by feature.

AMD has comparable tools, but the implementation story is less uniform. FSR spans multiple generations and technical methods, from spatial upscaling to temporal reconstruction and machine-learning models. Those differences protect broad compatibility, yet they also complicate feature expectations.

The mod therefore pressures AMD on software reach rather than simple feature count. Radeon already has an official answer for upscaling, generated frames, and reconstructed ray-traced detail. It does not yet offer a direct counterpart to DLSS 5’s generative final-rendering stage.

Nvidia’s Neural Rendering research describes a one-step, pixel-space diffusion model conditioned on the current frame, motion vectors, temporal state, and developer controls. The company says it preserves authored structure while adding learned appearance detail.

That is a different objective from reconstructing missing resolution. DLSS 5 can synthesize visual properties that the conventional renderer did not calculate in full. Nvidia positions it as a complement to physically based rendering rather than a substitute for game assets or artistic direction.

Blanco initially targeted that missing capability because no direct Radeon feature could enable it. The full suite arrived later, turning a solution for one absent function into a broader compatibility layer.

This creates an uncomfortable comparison for AMD. An independent developer is showing that Radeon hardware can execute workloads associated with Nvidia’s graphics stack, even if the results remain slower and less reliable.

The implication is not that AMD should adopt Nvidia’s technology. It is that Radeon users value flexible access across more games and hardware generations. The popularity of unofficial loaders reflects demand that official compatibility lists do not always satisfy.

The project had about 2,800 GitHub stars when reviewed on October 20. That figure is not proof of widespread daily use, but it signals meaningful attention for an alpha utility requiring manual libraries and unsupported game modifications.

For AMD, the forced response is long term. The company must expand Redstone support, improve cross-generation availability, and encourage developers to expose consistent inputs. Otherwise, community projects will keep defining what older Radeon hardware can attempt.

The Software Port Does Not Remove Nvidia’s Hardware Advantage

DLSS 5 on AMD reverses the access restriction, but it does not reverse Nvidia’s control of the models or its advantage in hardware tuned for neural workloads.

DLSS 5 officially targets GeForce RTX 50-series GPUs. Nvidia’s September DLSS 5 guidance says the technology works alongside Super Resolution, Frame Generation, Ray Reconstruction, ray tracing, and path tracing.

RTX hardware includes Tensor Cores dedicated to matrix operations used by neural networks. Radeon architectures can also run matrix math, but their resource allocation and supported formats differ. A software implementation must work within those architectural constraints.

Blanco’s runtime does exactly that by rebuilding the execution path for AMD GPUs. It maps the neural networks onto Radeon-friendly HIP kernels and memory layouts instead of relying on a native Nvidia runtime.

That is the mechanism behind the achievement and its limits. Portability comes from reimplementing execution, not from giving Radeon cards identical hardware. The closer a workload gets to sustained neural inference, the more those architectural differences can affect speed.

The project’s performance has improved rapidly. The 0.8.0 notes claim a 25 percent Neural Rendering improvement on RDNA 3. They also claim a 6 to 16 percent gain under heavy DirectX 12 gameplay loads from temporal accumulation changes introduced alongside version 0.7.0.

Those numbers describe developer-reported changes between builds. They are not standardized comparisons against native Nvidia execution. Game settings, output resolution, model selection, drivers, and rendering inputs can all change the result.

Blanco says an RX 9070 XT can reach 90 to 100 frames per second at 1080p while using the full HT Neural Rendering model. He also says output is similar to native DLSS, although some effects do not work completely.

The stated development goal is performance matching a GeForce RTX 5070 Ti. That remains an aspiration, not a verified result. No broad independent benchmark suite yet establishes parity across games or image-quality conditions.

Earlier tests illustrate how quickly performance can move and how difficult comparisons remain. September testing cited by DLSS-NR coverage showed Cyberpunk 2077 rising from roughly 30 frames per second to around 50 after rapid runtime improvements.

The same coverage described a claimed 74 percent improvement across two successive alpha builds. That gain was important for development, but it began from a low baseline. The game still ran substantially faster without Neural Rendering enabled.

That history suggests optimization headroom inside the runtime. It does not establish that software alone can close the remaining gap. Radeon execution may eventually reach a ceiling imposed by available matrix throughput, memory movement, or shared compute resources.

Frame Generation also complicates performance claims. Generated frames raise displayed frame rates without increasing the rate of new game-state updates. A higher counter can improve visual smoothness while input responsiveness remains tied to the underlying rendered rate.

Multi Frame Generation makes that distinction more important. If the system creates several frames around each traditionally rendered frame, artifacts and latency behavior depend heavily on the quality and timing of the base stream.

Ray Reconstruction has a different dependency. It relies on motion vectors, depth, surface information, and ray-traced inputs from the game. An incomplete or incompatible input path can reduce quality even when the neural model itself executes correctly.

The project’s modular design lets users combine inputs and outputs in unconventional ways. For example, a game’s XeSS inputs can feed Neural Rendering while FSR handles upscaling. That flexibility is valuable for experiments, but each combination creates another compatibility surface.

Nvidia retains an important advantage because it coordinates the original models, drivers, developer integrations, and hardware targets. It can tune the complete path instead of adapting to interfaces designed for another vendor.

AMD retains a different advantage through FSR’s broad deployment strategy. Some FSR versions work on several vendors’ GPUs, and developers can integrate them without requiring Radeon ownership. Redstone narrows that universality as AMD adds more machine-learning components.

Blanco’s project sits between those approaches. It treats vendor technologies as modules that can be detached from their intended hardware path. Its existence argues that the interfaces are more reusable than their marketing boundaries suggest.

The port does not make those boundaries meaningless. Official support still determines which configurations receive driver testing, developer assistance, predictable updates, and customer service. Mod compatibility can break when a game, driver, or vendor library changes.

The real reversal is therefore specific. Nvidia no longer has absolute technical exclusivity over where its visible DLSS outputs can appear. It still has the stronger claim to controlled performance and support.

Performance, Stability, and Anti-Cheat Keep the Win Qualified

The largest uncertainty is not whether DLSS functions can launch on Radeon, but whether ordinary players can depend on them across real games.

The project remains alpha software. Its 0.8.0 release notes list fixes for a GPU memory leak and a rare startup failure, while also identifying unresolved installation problems. The newer Epic Games launcher is not supported, and manually adding games can fail.

The stated requirements include Windows 11, an Adrenalin driver at version 26.1.1 or later, a DirectX 12 game, and user-supplied DLSS libraries. Vulkan and Linux support are planned but unavailable in the current release.

Anti-cheat presents a firmer boundary. The project warns that games with anti-cheat will block its DLL. Users must disable anti-cheat where a game permits that option, which generally limits experimentation to offline or otherwise approved modes.

That restriction removes many competitive multiplayer titles from the practical audience. It also reflects a sensible security design: anti-cheat systems are expected to reject unrecognized code injected into a game process.

Installation has improved through NeuralHub, but easier installation is not equivalent to consumer-ready behavior. A manager can centralize files and automate detection while underlying game compatibility remains fragile.

Independent testing has already demonstrated that fragility. One RX 9070 XT test encountered a Windows security warning, a black 3D scene, crashes, driver rollback, and problems launching Cyberpunk 2077 through Steam.

That test eventually ran Neural Rendering, but performance fell from 61 frames per second to 9 at 4K with ray tracing and balanced FSR. The result represented an 85 percent reduction in that configuration.

One trial cannot define every current build or configuration. It preceded some subsequent optimization, and Neural Rendering performance changes substantially with resolution, model settings, scene complexity, and pass count.

It nevertheless offers a necessary counterweight to the developer’s 1080p result. A feature that reaches 90 to 100 frames per second in one setup can still become impractical under a heavier workload.

Tom’s Hardware reported similarly mixed user results after version 0.8.0. Some users described little or no loss relative to native FSR, while others reported single-digit reductions or losses reaching 25 percent across different games.

Those accounts lack the controls of a formal benchmark suite. They still show that “works on Radeon” covers a wide range of experiences, from useful deployment to unstable experimentation.

Image quality needs the same caution. Blanco says output resembles native DLSS, except for effects that do not function completely. That qualifier matters because Neural Rendering changes materials and lighting rather than merely increasing sharpness.

A missing reflection, unstable face, altered surface, or temporally inconsistent detail can be harder to quantify than average frame rate. Side-by-side screenshots may hide problems that become visible in motion.

Multipass mode adds another tradeoff. Version 0.8.0 can run up to three Neural Rendering passes per frame, with separate styles and intensities. More passes can increase the visual effect, but they also add compute work.

The project also depends on proprietary model libraries supplied by supported games. Although Blanco says his code includes no Nvidia implementation, users still need compatible Nvidia DLLs. A future library update could change behavior or introduce a new compatibility challenge.

There is also no guarantee that game studios will support bugs caused by the loader. A developer cannot reasonably validate every combination of injected runtime, model file, upscaler input, and Radeon driver.

None of these cautions erase the technical accomplishment. Alpha modding projects often begin with narrow compatibility, manual installation, and large performance costs. Rapid iteration can solve problems that initially look fundamental.

However, the burden of proof has now changed. The first release proved that DLSS 5 Neural Rendering could execute on Radeon. Version 0.8.0 proves that a broader collection can share one modular framework.

The next proof must come from repeatable results. Reviewers need matching scenes, drivers, resolutions, model versions, and latency measurements across RDNA 2, RDNA 3, RDNA 4, and native GeForce hardware.

Until those results arrive, claims of native Nvidia quality or RTX 5070 Ti performance should remain attributed to the developer. The project has demonstrated access, not parity.

Three Signals Will Decide Whether the Mod Changes Radeon Gaming

The next stage will be determined by repeatable benchmarks, complete RDNA 2 support, and official responses from AMD or Nvidia.

The first signal is an independent benchmark set for version 0.8.0 or later. It should compare the same game scenes on an RX 9070 XT and an appropriate GeForce card while separating Super Resolution, Ray Reconstruction, Frame Generation, and Neural Rendering.

Average frame rate alone will not be sufficient. Testing should include low-percentile performance, input latency, frame pacing, video memory use, temporal stability, and visual artifacts. Generated frames must be identified separately from traditionally rendered frames.

Consistent results near the developer’s 1080p claims would strengthen the case that the runtime has crossed from demonstration into practical use. Large variations between titles would suggest that compatibility work remains more important than raw optimization.

The second signal is the project’s promised expansion for Radeon RX 6000 cards. Current documentation says RDNA 2 can run Neural Rendering under testing, while several other features are still missing.

Delivering most of the suite on RDNA 2 would show that the architecture abstraction scales beyond newer Radeon designs. It would also give a large installed base access to features unlikely to arrive through AMD’s newest official stack soon.

Failure to extend those modules would weaken the “three generations” framing. It would leave the complete suite concentrated on RDNA 3 and RDNA 4, with RX 6000 functioning as a limited experimental target.

The third signal is a response from the GPU vendors. AMD could accelerate Redstone availability, broaden older-hardware support, or improve driver-level upgrade paths for games with compatible FSR integrations.

Nvidia could leave the project alone, change its library behavior, or clarify how its model files may be used outside supported GeForce configurations. Even a technical update that unintentionally breaks the mod would reveal how dependent the project remains on Nvidia’s release cycle.

Game developers form part of this signal too. More standardized motion, depth, exposure, and ray-tracing inputs would make reconstruction technologies easier to exchange. Tighter validation or anti-tamper systems would move in the opposite direction.

The most likely near-term outcome is neither complete parity nor immediate failure. DLSS-NR-on-AMD can become a useful enthusiast tool while remaining unsuitable for buyers who expect one-click reliability.

That would still matter. Modders have repeatedly expanded the useful life of graphics features by creating wrappers, replacing libraries, and exposing options that vendors restrict by product generation.

Version 0.8.0 goes further by placing four headline technologies behind one Radeon-oriented manager. It shows that parts of modern graphics stacks can be separated from their intended vendor path, even when the final experience still depends on hardware.

For Radeon owners, the practical question is straightforward: would you trade official support and predictable compatibility for access to rendering features your card was never meant to run? Watch the next benchmark releases, the RDNA 2 feature list, and vendor updates before treating DLSS 5 on AMD as more than a compelling experiment.