NVIDIA DLSS 5 Mod Uses A Second GPU To Nearly Erase Its Performance Hit

NVIDIA's DLSS 5 neural rendering offers a stunning upgrade in visual realism but the performance hit is substantial, to the point that some, including this writer, feel it doesn't justify the cost. What if you could slash that performance hit down to almost nothing, though? Well, it turns out that's quite possible if you just run DLSS 5 on a second graphics card.

Hardware enthusiasts will likely recall that one of the early criticisms of the technology when it was first shown was that it required a pair of extremely expensive GeForce RTX 5090 graphics cards. The company has done an incredible amount of optimization on the technology since then, and it can now run on a single GeForce RTX 5060 alongside the game itself. That's undeniably impressive, but the AI model doing the neural processing remains very heavy. It causes every RTX GPU to slam against its power limit, drastically reducing performance. After all, the GPU still has to render the base game to give the DLSS 5 model something to work from.

However, as I discussed at length in my examination of the technology last week, DLSS 5 is essentially a post-processing filter. It arrives at the end of the pipeline to process the fully finished frame output by the game engine, ideally with the help of some engine-provided data although that's not necessary. Because it works this way, it is entirely possible to take the engine's output and feed it into a second graphics card dedicated to running the DLSS 5 model. This drastically increases the resources available both to the game engine and to the neural rendering process, allowing you to run DLSS 5 with a minimal performance hit.


How minimal? The developer of the ReShade plugin used to perform this process claims the performance drop ranges from 0% to 17%, depending on how hard the baseline GPU is working. Compared to running DLSS 5 on the same Geforce RTX 5060 Ti, where he's seeing performance hits between 36% and 59%, this is a relative uplift of up to 86%. If you increase the DLSS preset to Ultra Performance, reducing the load on the primary rendering card, you end up waiting on the second card to finish running DLSS 5.

However, there are a few caveats. This modification, known as MGPU Bridge, is based on ReShade and it requires addon support, meaning you won't want to use it in online games with restrictive anti-cheat software. It also currently requires at least one GeForce RTX 50-series card on which to run DLSS 5, though that requirement may soon be superseded given that developers have already gotten DLSS 5 running (albeit slowly) on RTX 40-series hardware. Finally, the developer recommends connecting a monitor to each graphics card, as running the secondary card headless is meaningfully slower; on his development machine, running headless cost "33% throughput and roughly doubled latency."


Interestingly, this method relies on ReShade without using the injection techniques seen in previous DLSS 5 mods. According to the developer, it does not feed motion vectors or depth data to the model; in this mod, the DLSS 5 model operates strictly off color data—that is, the finished output frame. In this setup, it functions as a pure image-to-image AI, much like uploading an image to a chatbot and asking it for edits, except executed 60 or more times per second. Consequently, the final output doesn't look quite as clean as a native DLSS 5 integration or even a mod that supplies depth and motion vectors; there's noticeable swimming in the shadows and fine details aren't as consistent as we saw in NBA 2K27.

If you're keen to try it yourself, head over to the maohgad-web GitHub repository for the Neural Coprocessor project and review the Readme file. Be prepared for a considerable amount of setup, as due to copyright restrictions, not all required files are bundled with the project. You'll need to source certain NVIDIA files independently and manually place them in directories.