DLSS 5 on Second GPU: This ReShade Bridge Offloads Neural Rendering to a Spare Card

By Indie Kings | September 13, 2026

Updated September 13, 2026: A new ReShade add-on called MGPU Bridge has crossed 100 stars on GitHub by asking a question PC gamers have been asking for years: what if a second GPU handled the AI upscaling while the first one renders the game? The project, which the author describes as research code with published measurements rather than a product, runs DLSS neural rendering on a spare graphics card and leaves the render GPU free. Measured on dual RTX 5060 Ti cards, it restored 86 percent of the performance headroom from upscaling, against 39 percent when the render card did the work. Here is how it works, the numbers, and the catch.

The Idea: Neural Rendering as a Terminal Stage
How MGPU Bridge Works
The Numbers: Dual RTX 5060 Ti
Requirements and Setup
What It Cannot Do Yet
Measured Results at a Glance
Bottom Line
Frequently Asked Questions

1. The Idea: Neural Rendering as a Terminal Stage

The premise behind the Neural Coprocessor project is simple, and it is stated plainly in the README: DLSS upscaling and neural rendering consume a finished frame and return a finished frame, so nothing about the work requires it to happen on the GPU that drew the picture. The add-on is named MGPU Bridge, it is a ReShade add-on loaded into a DirectX 12 game, and every log line it writes is prefixed [MGPU] in ReShade.log. The author is explicit that this is not SLI: SLI splits frame rendering across cards mid-frame, while the bridge moves a post-processing stage to another device entirely. It is also explicitly not a driver, not a patch, and not a replacement for anything, and it needs an add-on-enabled ReShade build (6.8.0 or newer) to load at all.

2. How MGPU Bridge Works

Under the hood, the bridge identifies the adapter the game renders on from the swapchain, creates its own Direct3D 12 device on a different adapter, copies each finished frame across a cross-adapter shared heap with a 64-byte seal per ring slot for frame identity and ordering, runs DLSS neural rendering on the second adapter once or twice per frame, and presents the result in its own window on the second card's monitor. The game's own rendering is never modified; the bridge reads finished frames and executes work elsewhere. Version 0.2.0 added two meaningful upgrades: the game's real depth buffer and motion vectors now reach the second card, giving neural rendering ground truth instead of motion derived from color, which reduces jitter and sizzle on faces and fine geometry, and both cards now have their own DLSS Super Resolution pass, so a weaker second card can do its neural work at a smaller resolution and let DLSS scale it back up to full. That, plus a move that puts nvngx_dlssnr.dll in a folder called mgpu instead of beside the executable, is what makes the technique work on older titles that would otherwise try to load the DLL themselves.

3. The Numbers: Dual RTX 5060 Ti

The measurements, published in the repository's RESULTS.md and summarized in the README, were taken in The Blood of Dawnwalker at 1920 x 1080 on two RTX 5060 Ti 16 GB cards, across the whole DLSS range. With DLSS 5 off, the game ran at 67 to 70 FPS in DLAA mode and climbed to 172 FPS at Ultra Performance. Turning DLSS 5 on and running it on the render card cut performance badly in every mode, down to 44 FPS in DLAA and 69 to 71 at Ultra Performance. Running the same DLSS 5 stage on the second card restored almost all of it: 67 to 70 FPS in DLAA, 106 to 107 at Performance, and 157 at Ultra Performance.

The author's headline finding is an efficiency comparison. Moving neural post-processing to a second GPU delivers 129 percent of available upscaling gains, which is 86 percent of the theoretical ceiling, while running DLSS 5 on the render card delivers only 59 percent, or 39 percent of the ceiling. In cost terms, offloading DLSS removed 36 to 59 percent of the workload from the render card depending on the mode, with almost no cost left on the second card at DLAA. The author notes the measurements are a floor, not a ceiling, since the implementation is crude and sits outside the engine.

4. Requirements and Setup

The requirements list is strict, and the README is honest about it. The primary path needs an RTX 50-series card (RTX 40-series is confirmed working with a modded nvngx_dlssnr.dll), two GPUs and two monitors with one card driving each, an add-on-enabled ReShade build 6.8.0 or newer, and a DirectX 12 game, since D3D11 and Vulkan are unsupported. No shader packs are required, and no other ReShade add-ons should be present to avoid multiple NGX consumers. The setup is three files: the add-on named so it contains nvngx.dll, the nvngx_dlssnr.dll inside an mgpu folder next to the add-on, and a mgpu_depth_tap.fx shader in ReShade's Shaders folder, which the add-on enables itself. In-game controls run the stream, toggle view modes (neural output, input, split), move the split seam, and adjust intensity, all through the ReShade overlay and hotkeys.

5. What It Cannot Do Yet

The limitations list matters as much as the win numbers. There is no resolution or DLSS mode change while armed, since that requires a swapchain rebuild. Frame generation is untested and not recommended. Color handling is not fully implemented, so the output can look washed out and the tone parameter has to be adjusted by hand. Motion vectors for UI and HUD elements are still missing. External overlays like RivaTuner and MSI Afterburner misbehave, so the project recommends ReShade's built-in FPS display. Single display setups are not supported in 0.2.0 for structural reasons: the second card presents through its own swapchain and window, and the author states that sharing glass with the game is a different architecture, not a setting. Photon-to-photon latency was not measured, and image quality was not assessed. The author flags all of this, plus the fact that the project is not affiliated with or endorsed by NVIDIA and carries a no warranty MIT license.

6. Measured Results at a Glance

Table 1: The Blood of Dawnwalker at 1080p, dual RTX 5060 Ti 16 GB (frames per second)

DLSS Mode	DLSS 5 Off	DLSS 5 on Render Card	DLSS 5 on Second Card
DLAA	67 to 70	44	67 to 70
Quality	98 to 99	54 to 55	91
Performance	127 to 131	59	106 to 107
Ultra Performance	172	69 to 71	157

Table 2: Where the workload landed and the efficiency headline

DLSS Mode	Cost on Render Card	Cost on Second Card
DLAA	-36%	0%
Quality	-45%	-8%
Performance	-54%	-17%
Ultra Performance	-59%	-9%

7. Bottom Line

The MGPU Bridge experiment is the most interesting DLSS project this year for anyone sitting on a spare GPU. The measured result, that offloading neural post-processing to a second card recovers 86 percent of upscaling headroom versus 39 percent on the render card, is exactly the kind of number hardware enthusiasts remember, and the method and mistakes are published openly in the repository. The catches are real: it needs a second GPU and monitor, an add-on ReShade build, and a DirectX 12 game, frame generation and latency are unaddressed, and the maintainer calls it research, not a product. For now it is a proof of concept that the GPU shortage era's most defensive question, can my second card do something useful, has a working answer.

Frequently Asked Questions

Can a second GPU run DLSS?
Yes, with this project. The Neural Coprocessor MGPU Bridge is a ReShade add-on that runs DLSS 5 neural rendering on a second GPU while the first renders the game, across a cross-adapter shared heap. It requires two GPUs, two monitors, and an add-on-enabled ReShade build.

Is this SLI?
No. SLI splits frame rendering across cards. The bridge works because neural post-processing is a terminal stage that consumes a finished frame and returns a finished frame, so it can run on another device entirely.

What GPUs work with MGPU Bridge?
GeForce RTX 50-series cards are the primary requirement for DLSS neural rendering, with RTX 40-series confirmed working using a modded nvngx_dlssnr.dll. The published measurements used two RTX 5060 Ti 16 GB cards.

What performance did the dual RTX 5060 Ti setup achieve?
In The Blood of Dawnwalker at 1080p, moving DLSS 5 to the second card restored performance to near the DLSS-off baseline in every mode, and the author reports it recovers 86 percent of the ceiling of upscaling gains versus 39 percent on the render card.

Does it work on any game?
Only DirectX 12 games, and only titles where the second card can share the finished frame. D3D11 and Vulkan are unsupported, single monitor setups are not supported in 0.2.0, and no other ReShade add-ons should be loaded at the same time.

Is MGPU Bridge safe to use?
It is research code under an MIT license with no warranty, not a product, and the author recommends reviewing code before execution. Known issues include washed-out color that needs manual tone adjustment, no frame generation, and external overlay tools misbehaving.

Table of Contents