Independent developer Marcelo Guibout has released "MGPU Bridge," a ReShade add-on that runs DLSS 5 neural rendering on a second GPU. It separates generative AI processing from the GPU that draws the game. In tests with two GeForce RTX 5060 Ti cards, it achieved game-side frame rates close to those measured with neural processing turned off.
However, the recovered numbers reflect how fast the game renders. They do not mean the processed image on screen refreshes at the same rate, nor that input lag disappears. What the published implementation and measurement records show is a benefit, getting rendering resources back, and a remaining problem in the display path. The developer's explanation makes this distinction clear.
Moving the post-render stage to another GPU
MGPU Bridge takes the color image a DirectX 12 game has finished rendering and passes it to a different GPU. The receiving GPU runs DLSS 5 neural processing and shows the result in its own separate window. The project is named "Neural Coprocessor." It is not an official NVIDIA feature but a research implementation published together with its measurement results.
The reason it can be separated lies in where the processing sits. Splitting the drawing of objects in a game across multiple GPUs requires coordinating data and processing order midway through rendering. What MGPU Bridge hands over is a single finished image from the game side. The destination does not take over the game's entire rendering; it handles only the final neural stage. According to the design document, images are passed through memory shared between the GPUs and displayed directly from the processing GPU, so results do not need to be returned to the original GPU.
It helps to separate DLSS 5 from conventional super resolution. Super resolution reconstructs a high-resolution image from one rendered at a lower internal resolution. DLSS 5 neural rendering uses learned knowledge of how the real world looks to generate lighting and material appearance. NVIDIA also describes it as an independent, optional feature that can be used alongside super resolution and frame generation.
For that reason, lowering the game's internal resolution does not lighten neural processing applied to the output-resolution image by the same proportion. When both run on the same GPU, neural processing consumes the headroom freed by lighter rendering, shrinking the FPS gain from super resolution. Moving the post-processing stage to another GPU aims to reduce this contention.
Which number should 157fps be compared with?
In the developer's comparison tests, the most closely matched conditions are in a setup called "Rig B." It uses the same location in The Blood of Dawnwalker with the camera held still and output resolution fixed at 1920×1080. It has a Ryzen 7 7800X3D and two RTX 5060 Ti 16 GB cards, each connected via CPU-direct PCIe 5.0 x8, with a monitor attached to each GPU.
The table below shows game-side FPS from Section 4b of the measurement document. The middle column adds neural processing on the same GPU via RenoDX through ReShade, and the right column offloads processing with MGPU Bridge.
| Rendering mode | Without neural processing | Neural on rendering GPU | Neural on second GPU |
|---|---|---|---|
| DLAA | 67–70fps | 44fps | 67–70fps |
| Quality | 98–99fps | 54–55fps | 91fps |
| Performance | 127–131fps | 59fps | 106–107fps |
| Ultra Performance | 172fps | 69–71fps | 157fps |
The left column is not a value with all DLSS features turned off. DLAA is an anti-aliasing method that renders at the same resolution as the output, and Quality and below use super resolution. The internal resolution for each row is the same across the three conditions; only the presence of neural processing and where it runs change.
In Ultra Performance, the game-side 157fps is roughly 121–128% higher than the 69–71fps with single-GPU processing, yet about 8.7% lower than the 172fps without neural processing. The former is calculated as 157 ÷ comparison FPS − 1, and the latter as 1 − 157 ÷ 172, both as percentages. Being much faster than running neural processing on the same GPU is a different judgment from returning completely to the state with processing off.
The "86%" figure in the developer's materials is also not an FPS improvement rate under identical settings. The FPS increase when switching from DLAA to Ultra Performance is about 151% without neural processing and about 129% with processing on a separate GPU. The 86% is a ratio showing how much of that increase was retained. It can be used to explain that the benefit of lowering resolution returns, but restating it as "86% faster with a second GPU" changes the denominator.
Some parts of the comparison are not matched. Processing intensity and number of passes are aligned, but RenoDX has a mask that narrows the processing target and MGPU Bridge does not. The timing at which images are received also differs. Furthermore, in the single-GPU condition the second GPU is disabled, so the system state is not completely identical. These numbers cannot be used as a performance comparison with an official DLSS 5 integration or with OptiScaler.
A fast game does not guarantee the display keeps up
MGPU Bridge's processed image appears in a window separate from the game's own screen. In the README, the developer states that the table shows the frame rate on the game-side monitor. Increasing the number of neural passes slows updates in that window, but the game-side FPS does not drop.
Because rendering and neural processing are separated, the two speeds can diverge. Even if the game keeps producing images, the destination cannot necessarily process them all at the same pace. MGPU Bridge also skips neural processing of stale images when newer frames have arrived. This design avoids accumulating delay by continuing to process old images, but it does not guarantee that every game-side frame is displayed after neural processing.
Therefore, even though DLAA's 67–70fps matches the processing-off value, it cannot be said that DLSS 5's load has become zero. The load has moved to another GPU. In the developer's environment, one neural pass on a 1080p image reportedly took about 8.3ms, but that is processing time on the second GPU. Total latency from input to the screen lighting up was not measured.
Monitor connection conditions also matter. According to the developer, in a configuration where no monitor is connected to the second GPU, throughput dropped 33% relative to the configuration with one connected, and measured latency roughly doubled. This figure reflects a difference from changing the display connection, not a result showing that going multi-GPU itself doubles input latency. Based on the README's explanation of display and latency, performance evaluation needs measurements of the neural screen's actual update rate and input response, in addition to game-side FPS.
Only the color image is transferred; motion is estimated at the destination
According to current development documents, only the color image crosses between GPUs, but the neural model is also given motion vectors. The destination GPU creates them using optical flow, which estimates motion from changes in the image. They are not received from the game engine. The README and the design document distinguish between these two stages.
This difference matters for image quality. The engine knows how objects and the camera moved, whereas estimating from a finished image requires finding correspondences from changes in visible pixels. Even if the amount of data transferred between GPUs is kept low, the processing does not use the same information as an official implementation.
In its DLSS 5 research overview published on September 1, NVIDIA says inference uses the current rendered frame and the engine's motion vectors, and also takes as input state carried over from past frames and control values for the look. It adopts a single-step diffusion model operating in pixel space, and during training it uses renderer-derived scene attributes to train for content consistency. The attributes used in training must be distinguished from the inputs always supplied while the game runs.
In the configuration MGPU Bridge tested, depth information is not passed to the model. This describes that configuration and is not proof that depth is unnecessary in every DLSS 5 implementation. The developer's performance measurements also exclude image-quality evaluation. To read the FPS difference as "the same image quality processed more cheaply," exposure would need to be matched and temporal stability compared in moving scenes.
What a two-GPU setup needs to become practical
The published FPS comparison covers one scene in one title, measured on two PCs, each with two RTX 5060 Ti 16 GB cards. There is also demonstration footage of Cyberpunk 2077, but it was not adopted in the measurement document, because after a settings change dropped frames and ordering problems were recorded and FPS was not reproducible. A working demo and performance evidence need to be viewed separately.
Stability testing is also limited to minutes. In the limitations section of the measurement document, the longest record of continuous normal transfer is 16,775 frames, about 5.5 minutes. It has not been confirmed that scene changes and settings changes during long play sessions are handled without problems. ReShade with add-on support is also required, and NVIDIA's DLLs for neural processing are not bundled with the project.
The economics of a two-GPU setup are also undetermined. There is no comparison against spending the same budget on a single higher-end GPU, and the effect of combining different GPUs cannot be predicted from this table. Even if the rendering card runs cooler, the second GPU still draws power and produces heat. Being able to distribute cooling is not the same as saving power across the whole PC.
Still, the implementation has shown that generative AI processing placed at the end of rendering can be detached from the device that draws the game. If future work compares final display FPS and input latency at equal image quality, along with long-session stability and whole-PC power consumption, it should become possible to identify when an additional GPU actually improves the play experience.