AMD Radeon Owners Can Now Run Nearly Every DLSS Feature Through Fan-Made Mod

For years, PC builders choosing AMD Radeon graphics cards had to accept a clear trade-off: higher raw rasterization performance and generous VRAM allocations in exchange for missing out on NVIDIA's proprietary Deep Learning Super Sampling feature set. Software modder Daniel Blanco has upended that paradigm with his DLSS-NR on AMD project, bridging the gap between competing GPU architectures through custom code. Can custom HIP kernels give your Radeon GPU full NVIDIA feature parity without introducing massive performance compromises?

The latest breakthrough moves this project far beyond a proof-of-concept hack into a multi-featured enhancement utility. By completely rearchitecting how graphics runtimes interact with non-NVIDIA silicon, custom software now allows select Radeon cards to execute complex neural upscaling and frame creation algorithms natively reserved for GeForce hardware.

What is the DLSS-NR mod, and how does it bring NVIDIA tech to AMD GPUs?

The DLSS-NR on AMD project is a community-developed software suite designed to execute NVIDIA's proprietary rendering algorithms on AMD hardware. Originally released in September as an experimental method to run DLSS 5 Neural Rendering on Radeon silicon, the project quickly demonstrated rapid development velocity when a single update unlocked a 74% performance jump in a single day.

The release of Alpha 0.8.0 on October 17 expands the software into a complete feature stack for Radeon owners. Instead of limiting users to basic upscaling, the mod enables DLSS 5 Neural Rendering, DLSS 4.5 Super Resolution using both Models M and L, Ray Reconstruction via Model F, and Frame Generation supporting up to 4x multi-frame generation (MFG) in supported titles, as originally reported by TechPowerUp.

To streamline deployment, the update bundles FSR 4.1.1 directly into the package so that AMD's native upscaler operates independently of driver updates. The release also introduces NeuralHub, a dedicated mod manager that handles automated game detection, DLL swapping, and configuration management across your installed library.

• DLSS 5 Neural Rendering: Real-time neural reconstruction replacing traditional spatial upscaling pipelines.

• DLSS 4.5 Super Resolution: Support for high-quality upscaling modes utilizing Models M and L.

• Ray Reconstruction: AI-driven ray tracing denoiser support running via Model F.

• Multi-Frame Generation: Optical flow frame insertion offering up to 4x multi-frame generation.

• Integrated FSR 4.1.1: Decoupled native upscaling capability free from Adrenalin driver updates.

• NeuralHub Management: Automated game detection and installation interface for quick mod deployment.

How does running DLSS via custom HIP kernels actually work without an NVIDIA card?

This is one of the most common misconceptions among system builders: the DLSS-NR on AMD project is not a simple translation wrapper or API hack. Traditional translation layers intercept API calls and map them onto equivalent functions, which often creates severe latency penalties and rendering bugs. Daniel Blanco built a ground-up reimplementation of the DLSS runtimes using custom Heterogeneous-compute Interface for Portability (HIP) kernels.

AMD's HIP environment allows developers to write C++ code that targets both NVIDIA CUDA hardware and AMD compute units. By writing targeted HIP kernels, the developer mapped the tensor matrix operations that DLSS expects directly onto AMD's vector and matrix compute pipelines. This enables Radeon cards to process the mathematical workloads of DLSS without needing dedicated NVIDIA Tensor Cores.

Because the mod acts as a runtime execution engine rather than a distribution of protected code, users must manually supply official, native NVIDIA DLL files from their own legal sources or local game installations. The mod engine hooks into these libraries at startup, redirecting their mathematical routines to the custom HIP kernels. On the software side, execution requires Windows 11 and games running exclusively on the DirectX 12 graphics API.

Custom HIP kernels bypass API translation overhead by compiling NVIDIA's neural routines directly for execution on AMD compute architectures.

Which AMD GPU generations support which DLSS features?

Feature execution varies significantly depending on the architectural capabilities of your AMD graphics card. While newer architectures feature dedicated AI matrix hardware that handles neural workloads, older generations lack the specialized compute structures required for heavy frame generation pipelines.

RDNA 4 (RX 9000 Series)

AMD's RDNA 4 architecture delivers full compatibility with every feature included in the DLSS-NR project. Hardware acceleration structures in cards like the RX 9070 XT allow smooth execution of DLSS 5 Neural Rendering, DLSS 4.5 Super Resolution, Ray Reconstruction, and multi-frame output up to 4x MFG.

Early testing demonstrates that an RX 9070 XT achieves roughly 90 to 100 FPS at 1080p resolution when running DLSS 5 Neural Rendering. The primary development objective for this architecture is matching the native feature performance of an RTX 5070 Ti.

Verdict: The ideal architectural tier for modded DLSS execution with full feature access.

RDNA 3 (RX 7000 Series)

The RDNA 3 architecture, powering cards like the RX 7900 XTX and RX 7800 XT, fully supports the entire modded feature stack. Dedicated AI accelerators present on RDNA 3 die designs handle the matrix math required for stable DLSS 4.5 Super Resolution and multi-frame frame generation without causing system crashes.

While rasterized performance is excellent, processing heavy neural models side-by-side with raster pipelines results in slightly higher latency compared to native GeForce hardware. Despite this overhead, frame-rate increases in heavy single-player games remain substantial.

Verdict: Excellent support tier providing access to all DLSS features and multi-frame outputs.

RDNA 2 (RX 6000 Series)

For owners of RDNA 2 graphics cards such as the RX 6800 XT or RX 6700 XT, software support remains strictly experimental. Because RDNA 2 lacks dedicated wave-matrix AI hardware instruction sets, the Alpha 0.8.0 release limits these GPUs exclusively to DLSS 5 Neural Rendering.

Advanced features like Ray Reconstruction Model F and multi-frame generation are entirely disabled or run at unplayable frame rates. System builders using RDNA 2 cards should treat the mod as a novelty rather than a daily rendering utility.

Verdict: Experimental tier restricted strictly to Neural Rendering with no frame generation support.

How does Radeon performance with DLSS compare to native FSR and NVIDIA cards?

Evaluating the real-world performance of an RX 9070 XT running modded DLSS requires comparing it directly against native FSR 4.1.1 execution and native GeForce hardware. While native FSR 4.1.1 delivers lower compute latency on Radeon hardware, the image reconstruction quality of modded DLSS 5 offers tighter temporal stability and reduced shimmering on thin geometric lines.

When running DLSS 5 Neural Rendering at 1080p, the RX 9070 XT operates in the range of 90 to 100 FPS, closely approaching the baseline performance target set by the RTX 5070 Ti. Scaling up to 1440p maintains consistent frame delivery, though total latency increases slightly due to custom HIP kernel context switching.

Frame-time delivery remains smooth across standard rasterization workloads, but introducing complex ray-tracing pipelines can lead to frame-pacing variances. For maximum frame consistency, native FSR 4.1.1 still holds a minor edge in baseline latency, but modded DLSS wins on image clarity.

While native FSR yields marginally lower frame latency, modded DLSS 5 provides superior temporal stability on AMD hardware.

Why do two Radeon GPUs with the same mod show wildly different Ray Reconstruction results?

Early testing across community forums reveals inconsistent performance trends when enabling Ray Reconstruction via Model F on Radeon hardware. In Cyberpunk 2077 utilizing path tracing at 1440p on an RX 9070 XT, one Reddit user reported zero performance penalty when toggling Ray Reconstruction on, whereas other testers documented frame rate drops of up to 25% under identical settings.

This performance divergence stems from how path-traced rendering pipelines distribute bounding volume hierarchy (BVH) traversal and denoising math across the GPU. NVIDIA Tensor Cores process spatial denoising, specular accumulation, and normal-map filtering simultaneously during ray reconstruction passes.

On AMD hardware, custom HIP kernels must pass those denoiser workloads through standard compute units. If a specific game scene is heavily bottlenecked by geometry or ray-triangle intersections, executing Model F denoisers on compute units steals resources directly from screen-space rasterization, causing performance penalties up to 25%.

• Scene Geometry Density: Complex environments increase BVH traversal strain, reducing compute capacity for software denoisers.

• Path Tracing Ray Counts: Higher per-pixel ray counts overload custom HIP kernel buffers, triggering performance dips.

• Compute Unit Contention: Software-based Ray Reconstruction competes directly with engine shading pipelines for shader engine registers.

• Memory Bandwidth Pressure: Accumulating spatial and temporal ray data demands high VRAM throughput, impacting high-resolution scaling.

What are the biggest misconceptions about running DLSS on AMD hardware?

With software modifications of this complexity, false information spreads quickly throughout builder communities. Addressing these urban legends head-on ensures you understand the exact security and technical risks involved before modding your system.

This is one of the most common misconceptions: believing that running modded DLSS on AMD hardware operates like a standard, risk-free graphics driver driver update. The reality is that hooking custom runtime DLLs into game engines alters system memory structures in ways that automated security tools monitor closely.

• Misconception 1: "Modded DLSS is just an API translation wrapper."
  Fact: The mod relies on a ground-up reimplementation of DLSS runtimes using custom HIP kernels designed to execute directly on AMD CUs.

• Misconception 2: "You can safely use this mod in online competitive multiplayer games."
  Fact: Anti-cheat software routinely detects hooked DLL files and custom runtime injection as memory manipulation, resulting in account bans. Stick strictly to single-player titles.

• Misconception 3: "Installing this mod will void your factory AMD graphics card warranty."
  Fact: Modifying user-space game software or swapping system DLL files has no impact on your physical hardware warranty. Physical hardware damage is required to void manufacturer coverage.

• Misconception 4: "The mod software illegally packages proprietary NVIDIA code."
  Fact: The download package contains open-source HIP injection hooks and the NeuralHub utility. Users must independently source their own official NVIDIA DLL files.

• Misconception 5: "Bundled FSR 4.1.1 still requires official driver support from AMD."
  Fact: The Alpha 0.8.0 release decouples FSR 4.1.1 from driver execution, allowing it to function independently inside any supported game directory.

Hooking custom runtime DLL files into game engines will trigger anti-cheat software, making the mod strictly intended for single-player games.

Where does modded DLSS fall short compared to native NVIDIA hardware execution?

While the achievement of running NVIDIA's feature suite on AMD hardware is impressive, user experience parity with native GeForce GPUs does not exist. System builders must evaluate several core operational trade-offs before choosing a software-bridged setup over native hardware execution.

Native GeForce hardware runs DLSS routines on fixed-function ASIC components (Tensor Cores and Optical Flow Accelerators) managed directly at the driver level. The modded Radeon implementation relies on software hooks, manually managed DLL dependencies, and custom context switching inside game engines.

Every official title update or game engine patch carries the risk of breaking hooked software functions, requiring you to wait for NeuralHub profile updates. Native NVIDIA hardware avoids these friction points entirely through official, driver-integrated implementation across the entire DLSS catalog.

Modded DLSS requires ongoing manual maintenance and limits game compatibility compared to native hardware options.

Does this mod change the price-to-performance math when buying a new GPU?

For budget-conscious builders, AMD cards generally provide higher raw rasterization performance per dollar and more generous VRAM capacities than competing NVIDIA models. However, buyers often opt for GeForce cards specifically to gain access to superior DLSS upscaling and frame generation.

The existence of DLSS-NR on AMD alters this value equation for tech-savvy users willing to manage manual software configurations. By buying a card like the RX 9070 XT or an RDNA 3 model like the RX 7900 XTX, you retain high native VRAM headroom while gaining access to DLSS 5 Neural Rendering and 4x multi-frame generation in single-player titles.

However, relying on community software modifications as a primary buying driver carries long-term risks. Developer Daniel Blanco creates these updates independently, meaning long-term maintenance, feature parity with future DLSS versions, and prompt game patch fixes are never guaranteed. If seamless execution across all games is your top priority, paying the premium for native GeForce hardware remains the safer investment path.

Conclusion: Should you buy Radeon now or stick to native NVIDIA hardware?

The DLSS-NR on AMD mod narrows the feature gap between competing GPU manufacturers, but your individual gaming habits should determine whether you rely on software bridging or buy native hardware.

If you play strictly single-player AAA titles and enjoy tweaking your system, buy a modern AMD Radeon GPU such as an RDNA 4 or RDNA 3 card. You will maximize your raw rasterization value and VRAM allocation per dollar while using the Alpha 0.8.0 release and NeuralHub to unlock DLSS 5 Neural Rendering and Ray Reconstruction in single-player experiences.

If you play competitive online multiplayer games, avoid this mod entirely and buy native NVIDIA hardware or stick strictly to official, driver-supported FSR. Multiplayer anti-cheat systems interpret hooked runtime DLL files as malicious memory injection, which will result in permanent account bans on competitive servers.

If you are a plug-and-play system builder who wants zero setup maintenance, buy native NVIDIA GeForce hardware like the RTX 5070 Ti. Native Tensor Core execution eliminates the need to manually source NVIDIA DLL files, monitor community updates, or troubleshoot game-patch incompatibilities.