A mod brings DLSS 5 neural rendering to a second GPU and almost eliminates the 50–60% performance drop

Developer Marcelo Guibout published on GitHub an add-on for ReShade called mgpu_bridge, which offloads DLSS 5 neural rendering workloads to a second graphics card in the system. This approach almost entirely eliminates the massive frame rate hit that comes with enabling DLSS Neural Rendering on a single GPU.

The idea isn't exactly surprising – the original DLSS 5 demos Jensen Huang showed off this spring ran on two graphics cards. The difference now is that a similar setup is available to enthusiasts, and an old GPU can stay in the case as a neural coprocessor after an upgrade.

Performance remains the biggest barrier for the technology. Nvidia itself points to a 50-60% frame rate drop as a baseline regardless of the game or how much AI filtering is happening, and NBA 2K27 tests on an RTX 5080 saw the hit reach as high as 73%.

The add-on works through ReShade, an app used to inject visual effects into PC games. The files can be downloaded from GitHub and added to a game that already has a DLSS 5 mod installed.

The setup works like this – the primary GPU renders the frame, and a second one (RTX 5060 Ti 16GB in testing) grabs the finished image, runs the neural network pass, and outputs the result to its own monitor. The frame never has to be handed back, which avoids extra overhead and latency.

The backstory behind the mod is interesting too. Guibout put together the initial add-on scaffold using a local Qwen3.8-27B-Q6_K model running in LM Studio on the same pair of RTX 5060 Ti 16GB cards, then wrote most of the C++ core through Claude, which, according to the developer, "also produced several confident, well-formatted, wrong diagnoses along the way."

The reason it is worth doing: neural post-processing saturates whatever device it runs on. 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. Move it to a second GPU and the render GPU is genuinely freed.

In the documentation, Guibout is upfront that the add-on isn't ready for everyday use and is published mainly for further research. Testing happened on exactly one machine across three games, and anyone interested is encouraged to check the source code first.

There's no shortage of limitations. A second monitor is required since the card that generates the neural output is also the one that displays it.

The second monitor is a requirement, not a nicety - the neural output is displayed by the card that produced it. Headless works and is meaningfully slower.

Cost is another factor. A pair of RTX 5060 Ti 16GB cards, sitting at the fifth tier of the Blackwell lineup, currently runs about $800 each, putting the total around $1,600 for the full setup.

The third point concerns the architecture of the solution itself. This ReShade implementation kicks in after the final frame has already been rendered, with no access to depth buffer or motion vector data that the official Nvidia version relies on. For a complete result, integrating the second card would require going much deeper into the rendering pipeline, though not as tightly as SLI once did.

There's a purely hardware-related headache too. With SLI and CrossFire gone, motherboards rarely offer multiple PCIe x16 slots these days, especially once NVMe drives start eating up available PCIe lanes.