A new mod for PC games aims to significantly reduce the performance impact of DLSS Neural Rendering by offloading its demanding AI computations to a separate graphics card. Developed as a ReShade add-on, this tool, named "mgpu_bridge," allows a secondary GPU to perform the neural rendering pass, preventing the primary rendering GPU from being saturated and thus mitigating the substantial frame rate penalties often associated with the technology. This approach mirrors Nvidia's original demonstration of DLSS Neural Rendering, which was shown running on two graphics cards.
Nvidia has previously indicated that enabling DLSS Neural Rendering can result in a 50% to 60% frame rate reduction, with some instances showing even higher impacts. This mod directly addresses that performance hit by freeing up the primary GPU. "Run it on the render GPU and lowering your render resolution just frees capacity the neural stage immediately eats, so upscaling stops paying for itself," developer Marcelo Guibout explained. "Move it to a second GPU and the render GPU is genuinely freed."
Guibout has made the ReShade add-on available on GitHub, noting that it is primarily intended for further research and development rather than immediate widespread use. The tool has been tested on a specific configuration involving two RTX 5060 Ti 16 GB cards. While the developer encourages users to review the source code, this development opens up possibilities for repurposing older GPUs as dedicated neural processing units in future systems.
However, the implementation comes with notable caveats. The most significant is the cost, as it requires two graphics cards. The tested setup utilized two RTX 5060 Ti cards, which together would represent a substantial investment. Furthermore, the current iteration necessitates the use of two monitors; the neural rendering output is displayed by the card that produced it, and running the system without a second display, while possible, results in slower performance.
Another limitation is that this ReShade implementation acts as a final step, processing the rendered frame after it has been completed. Unlike Nvidia's official implementation, it does not have access to depth or motion vector information, which are crucial for more integrated rendering processes. This means the second GPU must perform its tasks without the detailed context available to the primary rendering pipeline.
Despite these challenges, the mod represents a fascinating advancement in the DLSS 5 modding landscape. It offers a glimpse into how users might leverage existing hardware for new computational tasks, and it could potentially encourage Nvidia to integrate such dual-GPU functionality as a core feature to further solidify its ecosystem. The growth of community-driven development around DLSS Neural Rendering highlights the ongoing innovation within the PC gaming hardware space.