DLSS 5 Neural Renderer Runs on AMD Radeon GPUs via Mod

A community patch allows Nvidia's new neural rendering technology to run on AMD hardware, though it causes severe performance drops.
Key points
- A community mod allows Nvidia's DLSS 5 neural renderer to run on AMD Radeon GPUs by intercepting FSR data streams.
- DLSS 5 is a neural rendering technology that adds lighting and material detail, distinct from traditional upscaling or frame generation.
- Performance testing on an RX 9070 XT showed frame rates dropping from 90 FPS to under 20 FPS when the feature is enabled.
Nvidia's latest neural rendering technology, known as DLSS 5, is no longer exclusive to RTX graphics cards. A community-driven workaround allows users to run this proprietary software on AMD Radeon GPUs, including the RX 9070 XT. While the modification successfully executes the full neural rendering pipeline on non-Nvidia silicon, it comes with a significant performance penalty that limits its practical use.
This development highlights the ongoing tension between vendor-specific features and open hardware standards. The mod does not simply reskin AMD's existing upscaling technology; instead, it injects Nvidia's actual neural renderer into the game's rendering pipeline. For enthusiasts, this serves as a fascinating proof of concept, demonstrating that the underlying mathematics of the technology are not inherently tied to specific hardware architectures.
Neural rendering differs from upscaling
It is crucial to distinguish DLSS 5 from previous iterations of Nvidia's Deep Learning Super Sampling. While earlier versions focused on upscaling lower-resolution images or generating intermediate frames to boost frame rates, DLSS 5 operates as a neural renderer. It sits at the very end of the rendering pipeline, taking the final color output and motion vectors from a completed frame to add complex lighting, material details, and ambient occlusion.
According to XDA Developers, this technology is designed to be deterministic, ensuring that visual elements like character faces remain stable from frame to frame. Game developers can adjust two specific parameters: Structure Intensity, which controls high-frequency details like reflections, and Tone Intensity, which manages broader lighting and color responses. This approach fundamentally changes how light and texture are calculated, offering a different visual fidelity compared to traditional upscaling methods.
The FSR dependency enables the mod
The mechanism behind this workaround relies heavily on AMD's own FidelityFX Super Resolution (FSR) technology. For the mod to function, the game must already support FSR. This is because FSR generates the necessary color and motion vector data for every frame. The mod intercepts this data stream, which is precisely what Nvidia's neural renderer requires as input, effectively using AMD's upscaler as a data feeder for Nvidia's renderer.
Setting up the software requires Windows 11 and a recent version of AMD Adrenalin drivers. Users must extract the neural renderer DLL file from a compatible game, such as NBA 2K27, and place it into the target game's directory. Once installed, an overlay accessible via the End key allows users to toggle the feature on and off instantly. This strange hybrid pipeline, where AMD's tech feeds Nvidia's AI on AMD hardware, is only possible because of this data overlap.
Performance drops to unplayable levels
Despite the technical success of running the renderer on Radeon hardware, the performance cost is substantial. Testing in Cyberpunk 2077 on an RX 9070 XT showed that enabling DLSS 5 drops frame rates from a smooth 90 FPS at 4K resolution to below 20 FPS. While the visual result adds noticeable depth to character skin textures and lighting, the reduction in frame rate makes the experience unplayable for most users.
The trade-off is clear: users gain a higher level of visual fidelity in specific areas, such as material detail and lighting, but lose the fluid motion required for gaming. This performance hit is not unique to AMD cards, but the fact that it occurs on hardware that is not natively designed to run this specific Nvidia workload underscores the heavy computational load of the neural renderer. For now, this remains a niche experiment rather than a viable alternative to standard upscaling techniques.






