A developer has vibe-coded a new ReShade add-on file that offloads the DLSS Neural Rendering workloads to a second GPU, almost negating the otherwise massive DLSS NR performance hit. Given the original DLSS 5 shots we saw when Jen-Hsun first showed off the feature around GDC and GTC in the Spring were running on two graphics cards, that’s not wholly surprising, but this mod certainly does open the the potential for keeping your old GPU around for neural coprocessing in the future.
Like almost all of the current DLSS 5/Neural Rendering implementations we’ve seen where the user gets to pick the game, the model, and the weight of the effects, the Neural Coprocessor add-on file—mgpu_bridge—is designed for ReShade, an app used to inject different visual effects into PC games. The files exist on Github, can be downloaded and added into a game that’s already been modded with DLSS 5 support, and could dramatically improve the performance.
This is, for me, the biggest issue with both the homebrew mods for DLSS Neural Rendering and the only release implementation we’ve seen from Nvidia and the developers themselves: the hefty performance penalty from turning it on.
Nvidia itself has touted a 50 – 60% frame rate hit as a blanket figure—regardless of game and how much AI filtering it’s doing—as what it is expecting to see from enabling DLSS NR. Andy benchmarked NBA 2K27 ahead of launch and found that sometimes the frame rate impact can be even higher than that.
I mean, it’s not as bad as when you try and run DLSS NR on AMD GPUs, but I’ve noted these kinds of performance hits when you run the new feature through ReShade, too. But Marcelo Guibout has coded a ReShade add-on file that allows a second GPU—here an RTX 5060 Ti 16 GB—to take the final finished frame and perform the neural rendering pass on that GPU and output on its own monitor. That means it doesn’t have to hand the frame back, which adds its own performance hit and latency.
They initially used a local model—Qwen3.8-27B-Q6_K—running on LM Studio using the same pair of RTX 5060 Ti 16 GB cards they eventually tested the feature on. This created the “initial add-on scaffold” which they then took into Claude to write most of the C++ core, though Anthropic’s AI “also produced several confident, well-formatted, wrong diagnoses along the way.”
Looking through the Github documentation Guibout makes it clear this is not a production-ready add-on, and is mostly presented for further research and development. They note it has been run on exactly one machine with three games, and encourage everyone to read the source code before running the files. But it’s still a really interesting development in the DLSS 5 modding story.
“The reason it is worth doing: neural post-processing saturates whatever device it runs on,” writes Guibout. “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.”
DLSS NR on AMD
AMD GPUs running DLSS 5 looks great, but runs horribly.
DLSS NR Off
DLSS NR On
This is almost certainly the reason behind Nvidia originally running the feature on two GPUs for the original demos, and the fact that it can get it running on a single GPU is still impressive from a technical perspective, but having a coprocessor handling the load should give you the best of both worlds.
Though there are obvious caveats to this.
The first being that it’s expensive enough to buy one graphics card right now. You look at this mod and the fact it’s been created and tested on a pair of RTX 5060 Ti cards and think that’s kinda mid-range. They are the fifth-tier cards of Nvidia’s RTX Blackwell stack. But those are two GPUs that now cost something around $800 apiece, so we’re still talking about needing $1,600 worth of graphics silicon to get this working.
The other caveat is that we’re running on with two monitors here as a necessity. Guibout notes in the Github readme file that, “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.”
The third caveat here is that this ReShade implementation is a final step process. It is handled after the final frame has been rendered and works with none of the depth and motion vector information the official Nvidia implementation runs with. That means the second-GPU is going to have to be more tightly integrated into the actual rendering process. Not quite as tightly bound as with SLI, perhaps, but it will surely still need to get involved before the final post-processing is done.
But I could potentially see this as a way to do something useful with your current-gen card when you next upgrade. And, on a purely mercenary level, it might be a smart thing for Nvidia to actually make as a feature as it would have the effect of locking someone into the GeForce ecosystem even more securely.
(Image credit: Future)
Though times have changed and SLI and CrossFire have died, and so motherboards are even less likely today to have multiple PCIe x16 sockets available to stick secondary GPUs into. Especially if you’ve got multiple NVMe SSDs taking up precious PCIe lanes.
I’m loving the rate of development happening outside of Nvidia now that the feature has fallen into the hands of the community at large. And regardless of how you feel about the technology, I would encourage you, if you have the hardware, to take a look at how even just the ReShade implementation looks, and how you can alter it. It gives a fascinating insight into the technology, into its limitations as well as its unwelcome excesses.
Though it’s not always easy. I’ve spent the morning trying to get this ReShade add-on working with the test rig, because I was interested to see if it could be run asymmetrically, as I don’t have a pair of RTX 5060 Ti 16 GB cards. Sadly, with an RTX 5080 and RTX 5060 Ti as the neural coprocessor, I could not get the add-on to arm when in Cyberpunk 2077. So, I’ve got more work to do to get this thing running.
