Mod Offloads Nvidia DLSS 5 Neural Rendering to Second GPU to Boost FPS

Developer Marcelo Guibout has released a new mod that offloads Nvidia's DLSS 5 Neural Rendering (NR) to a secondary graphics card. By separating the AI workload from the primary rendering process, the tool prevents neural post-processing from bottlenecking the main GPU.

The mod utilizes a ReShade add-on called "MGPU Bridge" to create a D3D12 device on a second GPU. Under this configuration, the primary GPU handles the core game rendering, while the second GPU takes over the neural post-processing tasks and displays the final result. According to Guibout, this is possible because neural rendering occurs at the end of the frame pipeline, taking a finished frame and returning a finished frame, which allows the process to be shifted to separate hardware.

The Performance Impact

Nvidia's DLSS 5 Neural Rendering applies AI-driven post-processing to frames at the output resolution. Because this happens at the end of the pipeline, it can consume significant resources, often offsetting the performance gains typically achieved by lower-resolution upscaling on a single card.

In tests using "The Blood of Dawnwalker" at 1080p in Performance mode, the impact of offloading was substantial. When running on a single card, the system achieved 59 FPS. By offloading the neural rendering to a second GPU, performance jumped to between 106 and 107 FPS, recovering a significant portion of the native performance, which sits between 127 and 131 FPS. Beyond raw speed, the separation provided thermal benefits; the primary rendering GPU ran 21 degrees Celsius cooler when the neural load was moved to the second card.

A Return to Coprocessors

This approach mirrors the historical use of dedicated PhysX cards, where specific physics calculations were offloaded to a secondary processor to free up the main GPU. By treating the second card as a "neural coprocessor," the mod demonstrates that AI rendering can be decoupled from the main geometry and shading pipeline. This suggests a potential future for hardware configurations where a low-power AI accelerator handles the heavy lifting of neural post-processing while a high-performance GPU focuses on rendering.

Implementation and Trade-offs

Despite the performance gains, the setup involves specific hardware requirements and technical compromises. The system requires two monitors—one connected to each graphics card—to function. The project is currently available for download on GitHub under the name "Neural-coprocessor."

While the mod proves the viability of decoupled AI rendering, it remains a community-driven experiment. It will be worth watching if this "coprocessor" logic influences future official hardware designs or if Nvidia integrates similar multi-chip strategies to handle the increasing demands of neural rendering.