A technical breakthrough by developer Marcelo Guibout has demonstrated a way to decouple the heavy computational load of Nvidia's DLSS 5 from the primary rendering GPU. By utilizing a ReShade add-on called MGPU Bridge, Guibout has managed to offload the Neural Rendering (NR) stage to a second graphics card, effectively treating it as a dedicated AI processor for the final frame output.
The MGPU Bridge Mechanism
The core of this mod lies in the nature of DLSS 5's generative rendering. Because neural post-processing occurs at the very end of the frame pipeline—taking a finished image and returning a refined one—it can be intercepted. The MGPU Bridge creates its own D3D12 device on the second GPU, transmits the completed frame from the primary card, executes the DLSS-NR pass on the second device, and outputs the result directly to a monitor connected to that second card.
This architecture prevents the primary GPU from being saturated by AI tasks, allowing it to focus entirely on rasterization and ray tracing. In tests using two RTX 5060 Ti 16GB cards on a Ryzen 7 7800X3D system, the primary GPU ran 21 degrees cooler due to the absence of neural workloads.
Performance Gains and Trade-offs
While Guibout clarifies these are technical showcases rather than strict benchmarks, the data from The Blood of Dawnwalker demo at 1080p reveals a significant performance delta. When DLSS 5 is run on the primary card, it consumes an increasing share of each frame as internal resolution drops, limiting the gains from upscaling. By moving this load to a second GPU, Guibout reports retaining 86% of the machine's potential performance gain across various DLSS modes.
The most notable impact is seen in the "Ultra Performance" mode, where offloading helps maintain higher frame rates compared to single-GPU execution. However, this setup is not without costs: it requires a second display and results in the doubling of display latency.
A New Era for Modding DLSS 5
This development follows a wave of community efforts to expand the reach of Nvidia's latest AI tech. Recently, modders have successfully patched DLSS 5 for RTX 4000 series and even brought it to AMD RDNA 4 GPUs via API wrappers.
Unlike the old SLI technology, which split the rendering of a single frame across GPUs, this method mirrors the legacy of dedicated PhysX cards. It allows enthusiasts with multi-GPU setups to bypass the performance bottleneck created by high-resolution neural filters without needing a single, massive flagship card.
No comments yet. Be the first!