Gibson Chan/pexels.com

A community mod that launched six weeks ago as a narrow experiment has quietly become a complete platform. As of October 17, 2026, DLSS-NR on AMD — built from scratch by solo developer Daniel Blanco — delivers all four of Nvidia's core DLSS 5 technologies to Radeon GPUs spanning three generations of silicon, reaching back to the RX 6000 series released in 2020. Blanco's project has reached a full four-technology suite across three Radeon GPU generations in under seven weeks.

For anyone who has hesitated over an AMD GPU purchase because DLSS 5 Neural Rendering — the AI lighting and materials pass that has become Nvidia's most-discussed differentiator in 2026 — was exclusive to GeForce hardware: that argument is no longer clean-cut.

What v0.8.0 Adds

When Blanco published the project's first build on September 3, 2026, the same day Nvidia officially launched DLSS 5 with NBA 2K27, DLSS-NR on AMD did one thing: get the Neural Rendering feature running on RDNA 4 hardware. Performance was rough — Cyberpunk 2077 dropped from 80 frames per second (FPS) to around 11-12 FPS at 1080p on an RX 9070 XT in early testing. The project was technically significant but practically unplayable.

Version 0.8.0, released October 17, 2026, is a different proposition. The full DLSS 5 suite is now available across all four technologies:

DLSS Super Resolution 4.5 — the AI upscaler that reconstructs a high-resolution frame from a lower-resolution input using temporal history and motion vectors.

DLSS Frame Generation — multi-frame synthesis that generates up to four interpolated frames for each rendered frame, dramatically increasing apparent frame rates.

DLSS Ray Reconstruction — AI-based denoiser for ray-traced lighting, replacing stochastic noise in ray-traced scenes with inferred detail.

DLSS 5 Neural Rendering (build 310.9.16) — the flagship per-pixel AI pass that adds photorealistic lighting and material detail that the game's conventional renderer never calculated. This is the capability AMD has no official equivalent for.

The update also bundles FSR 4.1.1, giving the mod a consistent upscaling baseline regardless of which FSR version a given game ships with. A parallel open-source runtime, AMDNR, developed by a separate contributor known as Kien (GitHub handle lmxxf), is selectable from the same launcher and provides similar functionality with a fully auditable codebase. Both runtimes use custom HIP kernels on Radeon GPUs rather than translating Nvidia GPU code at runtime.

How Does DLSS Run on Non-Nvidia Hardware?

The common assumption — that DLSS is inseparable from Nvidia's GPU architecture — is half right. Nvidia's DLSS 5 Neural Rendering model is trained and packaged as a set of weights inside a 158-megabyte (MB) dynamic link library (DLL), nvngx_dlssnr.dll. The weights encode a 71-block Swin/ViT neural network. What the weights do not encode is the execution engine — the code that actually runs the network inference on specific hardware.

Blanco's mod supplies that execution engine as a set of custom HIP compute kernels — AMD's GPU compute programming model, structurally similar to Nvidia's CUDA — compiled for specific RDNA architecture targets. The game's upscaler pipeline is intercepted; the mod inserts the neural rendering pass using the weight data from the user-supplied Nvidia DLL, but runs inference entirely on its own HIP kernels. Nvidia's runtime code never executes on AMD hardware.

This is a meaningful technical distinction. The mod is not a translation layer or a compatibility shim — it is a from-scratch reimplementation of the neural network runtime, written specifically for AMD GPU architectures. That distinction also provides some legal insulation: the mod does not contain or redistribute Nvidia's code. Users must supply the Nvidia DLL from a game, typically obtained from a legally purchased game that ships DLSS 5 support, such as NBA 2K27.

Why Does FP8 Make This Possible on RDNA 4?

The shared technical foundation that makes the port viable is FP8 (8-bit floating point) arithmetic. DLSS 5's neural inference pipeline is designed around FP8 precision — a compact numerical format that dramatically reduces memory bandwidth and compute requirements compared to the FP16 or FP32 formats used in earlier DLSS generations. Nvidia's RTX 40-series (Ada Lovelace) and RTX 50-series (Blackwell) GPUs support FP8 natively via Tensor Cores. AMD's RDNA 4 architecture natively supports FP8 and INT8, via dual AI accelerators per Compute Unit with sparsity acceleration.

That shared hardware capability is what the mod exploits. Where Nvidia's implementation uses Tensor Cores — hardware units purpose-built for matrix math — the mod runs the same FP8 operations on RDNA 4's AI accelerators.

The GPU generation gap in the mod's support table is explained by the same math. RDNA 3 (RX 7000 series) lacks native FP8 support, so the mod must use software fallbacks or FP16 emulation, which carries a performance penalty. RDNA 2 (RX 6000 series) has no native FP8 at all, which is why version 0.8.0 supports only the Neural Rendering pass on those older cards — the most minimal workload subset — with additional features planned for future releases.

How Does It Install?

Installation no longer requires manually copying DLL files into game directories. Version 0.8.0 ships through NeuralHub, a mod loader that presents a graphical user interface (GUI), scans for compatible DirectX 12 titles, and injects the mod into whichever upscaling pipeline a given game exposes — FSR, DLSS, or XeSS passthrough are all supported as injection points. An in-game overlay lets users tune Neural Rendering intensity, local tone behavior, and skin structure without leaving the session.

Users still need to supply the Nvidia DLL and must be running Windows 11 with AMD Adrenalin driver version 26.1.1 or later. The mod does not work in games protected by active anti-cheat systems — Easy Anti-Cheat, BattlEye, and Valve Anti-Cheat all detect and block the DLL injection technique the mod relies on. This is a structural limit, not a bug: DLL injection is how cheating tools operate, and anti-cheat systems are designed to prevent it categorically. The practical effect is that the mod is limited to single-player games and multiplayer titles that either do not use anti-cheat or allow it to be disabled.

Does It Actually Work? What Are Real-World Performance Results?

The performance story is more nuanced than the headline numbers suggest. Three releases between late September and early October 2026 produced substantial gains: Alpha 0.3.3 delivered a 23% improvement over 0.3.2; Alpha 0.4.0 added a further 42%; Alpha 0.4.1 added another 8% — a combined improvement of roughly 89% across those three releases in under two weeks. Tweaktown confirmed the 89% total performance gain across all three alpha updates.

In the developer's own testing by version 0.8.0, DLSS-NR delivers 90–100 FPS at 1080p on an RX 9070 XT. Community results are more varied: some users on Reddit report no meaningful frame-rate penalty compared to native FSR, while others document losses ranging from a few FPS to about 25%, depending on game, resolution, and system configuration.

The gap between developer testing and community testing reflects a genuine engineering challenge. Nvidia's Blackwell Tensor Cores were designed with substantially higher FP8 throughput than AMD's RDNA 4 AI accelerators. Running a 71-block transformer network per frame, at interactive frame rates, pushes both architectures — but Nvidia's hardware was purpose-built for it. Blanco's stated long-term target is matching RTX 5070 Ti per-frame performance on an RX 9070 XT. He describes making improvements daily.

Does This Actually Change Anything for AMD GPU Buyers?

The straightforward answer is: somewhat, for some buyers, right now.

Until October 2026, DLSS 5 — particularly Neural Rendering, the feature AMD has no official equivalent for — was a real and commonly cited reason to choose Nvidia over AMD even when AMD's raw rasterization performance was competitive. The RX 9070 XT had been widely assessed as strong competition for the RTX 5070 on conventional benchmarks; the software-feature gap was the counterargument.

That counterargument is now more complicated. The mod is not a full replacement for native DLSS 5 on Nvidia hardware: the performance gap is real, setup requires more steps than simply enabling a driver feature, anti-cheat incompatibility eliminates a large portion of the game library, and performance varies enough across configurations that some users may see meaningful frame-rate penalties. These are real costs.

At the same time, the mod demonstrates something editorially important: DLSS 5 Neural Rendering's vendor exclusivity is a software policy, not a hardware necessity. AMD's RDNA 4 architecture supports the same FP8 arithmetic that DLSS 5 requires, and a solo developer using AMD's own open HIP/ROCm compute stack replicated the neural network runtime from scratch in six weeks. Buyers who are comfortable with the enthusiast-modding ecosystem and primarily play single-player titles now have a viable path to DLSS 5 Neural Rendering on AMD hardware.

Is AMD Building Its Own Answer?

Yes, reportedly — though not for current hardware. Known hardware leaker Kepler_L2 posted on the AnandTech forums in September 2026 suggesting AMD is developing "Neural Lighting" for its next-generation RDNA 5 architecture, expected with the RX 10000 series. AMD's own published research in 2026 included work on real-time neural indirect illumination — one presentation at SIGGRAPH 2026 and a separate generative illumination paper published for ECCV 2026 — though neither paper names a product or a launch date.

AMD has not officially announced Neural Lighting or any DLSS 5 equivalent, and RDNA 5 has no confirmed launch date as of this writing. The reported approach — more deeply integrated into game engines than DLSS 5's general screenspace model — would likely require new hardware to run at acceptable frame rates, meaning RDNA 4 owners buying today may be looking at the community mod as their most realistic near-term path to this class of enhancement.

Neither Nvidia nor AMD has commented publicly on the DLSS-NR on AMD project, despite the mod receiving coverage across major hardware outlets. No cease-and-desist activity has been reported.

What Still Needs to Work

Blanco has signaled several near-term priorities: extending the full DLSS feature set to RDNA 2 hardware (currently limited to Neural Rendering), continued HIP kernel optimization to close the throughput gap with native Nvidia performance, and Vulkan API support alongside the current DirectX 12 pipeline. The parallel AMDNR runtime by Kien has expanded to support handheld AMD APUs including the Z1 Extreme and Radeon 890M-class hardware, suggesting the project's scope is still widening.

For a project that went from a proof-of-concept to a four-feature platform in under seven weeks, the optimization trajectory has been faster than most observers expected. The gap between "technically interesting" and "practical daily driver" has not closed yet — but the rate at which it is closing suggests that measurement will look different again in another six weeks.

Frequently Asked Questions

Does DLSS-NR on AMD work on all AMD Radeon GPUs?

Version 0.8.0 supports three generations of Radeon hardware: RDNA 4 (RX 9000 series) receives the full DLSS 5 suite including Neural Rendering, Super Resolution 4.5, Frame Generation, and Ray Reconstruction. RDNA 3 (RX 7000 series) also receives the full feature set, though performance is lower because RDNA 3 lacks native FP8 (8-bit floating point) arithmetic support. RDNA 2 (RX 6000 series) currently supports only the Neural Rendering component, with additional features listed as planned. The mod requires Windows 11 and AMD Adrenalin driver 26.1.1 or newer.

If the mod is a "reimplementation," how is it different from just running Nvidia's code on AMD hardware?

The mod does not run any Nvidia code. What it does use is the set of trained neural network weights contained in Nvidia's nvngx_dlssnr.dll file — the mathematical model that describes how the DLSS 5 neural rendering network behaves. What the mod supplies independently is the execution engine: custom HIP compute kernels written for AMD GPU architectures that actually run the network inference. In this model, Nvidia's file provides the "what" (the trained AI brain), and Blanco's mod provides the "how" (the hardware-specific code that runs it). Users must supply the Nvidia DLL themselves, typically from a legally purchased game that ships DLSS 5 support.

Can someone use this in online multiplayer games?

No. The mod relies on DLL injection — inserting code into a running game process — which anti-cheat systems including Easy Anti-Cheat, BattlEye, and Valve Anti-Cheat are specifically designed to detect and block. This is a fundamental structural incompatibility, not a temporary limitation: the same technical method that cheating tools use to inject aim-assist or wallhack code is the method the mod uses to insert its neural rendering pass. Anti-cheat systems cannot distinguish intent. The mod is limited to single-player titles and multiplayer games that do not use active anti-cheat enforcement.

When will AMD release its own official version of DLSS 5 Neural Rendering?

AMD has not announced any official equivalent. Leaker Kepler_L2 reported in September 2026 that AMD is developing a feature internally called "Neural Lighting" for its next-generation RDNA 5 architecture, but RDNA 5 has no confirmed launch date, and AMD has not publicly confirmed the feature, its specifications, or its hardware requirements. AMD's published research in 2026 covers real-time neural indirect illumination — one SIGGRAPH 2026 presentation and one ECCV 2026 paper — which is related but is not a product announcement. RDNA 4 owners purchasing today should not assume they will receive an official AMD neural rendering feature in a future driver update.