A ReShade Mod Runs DLSS 5 on a Second GPU

MGPU Bridge hands the DLSS 5 neural pass to a second graphics card and gives the render GPU most of its frame rate back, while separate mods have the same runtime loading on Radeon and 4x frame generation running on Ada.

Saya5 minute readDrivers & Software

DLSS 5 Neural Rendering is a post-process pass: it takes a finished frame and hands a finished frame back. That property turns out to be exploitable, and a developer has used it to move the whole pass onto a second graphics card while the first one keeps rendering the game. Two other mods from the same week push the same runtime onto Radeon and unlock 3x and 4x frame generation on RTX 40.

MGPU Bridge moves the neural pass to a second card

The project is called Neural Coprocessor, and the piece a user installs is a ReShade add-on named MGPU Bridge, whose first precompiled build, 0.1.0, went out on September 7. Marcelo Guibout demonstrated it with two RTX 5060 Ti 16GB cards on a Ryzen 7 7800X3D, both on CPU-attached PCIe 5.0 x8 links and each driving its own monitor. The add-on creates a D3D12 device on the second GPU, copies each completed frame across, runs the neural model there, and displays the result on that card's own display, so nothing travels back to the render GPU.

The numbers come from a static scene in The Blood of Dawnwalker at 1080p, with DLSS Super Resolution active in every column and only the neural pass moving. In Performance mode the game runs at 127 to 131 FPS with Neural Rendering off, 59 FPS with it on the render card, and 106 to 107 FPS with it on the second card. Ultra Performance goes 172, then 69 to 71, then 157. Guibout's reading of his own table is the useful part: the neural pass saturates whichever device it lands on and always runs at output resolution, so its cost barely falls as the DLSS mode drops while the render work collapses, which is why upscaling stops paying for itself on a single card. Offloading keeps about 86 percent of the available gain against roughly a third when one card does both, as Tom's Hardware reports with the full table. The render card also ran 21 degrees cooler without the neural load.

What the table measures is the game's frame rate on the render GPU, not the rate of finished neural frames reaching the second monitor, and ComputerBase makes the distinction explicit: the two run at separate rates, the bridge skips frames the post-processing has fallen behind on, and no end-to-end input-to-photon measurement has been published. Tom's Hardware puts the price as doubled display latency plus the second display. Guibout is clear that this is not SLI, since the game stays entirely on one GPU and the split falls between rendering and post-processing, the same trick Lossless Scaling users already run for frame generation, as VideoCardz notes.

The setup is early: it wants two RTX 50 cards, one monitor each, ReShade 6.8.0 or newer with add-on support, a DirectX 12 game and NVIDIA's nvngx_dlssnr.dll; frame generation is untested; and changing resolution or settings mid-run can break it. The comparison is also mod against mod, RenoDX on one card against MGPU Bridge on two, so it does not transfer to a native integration like NBA 2K27's. Nobody should buy a second GPU for this before someone measures the latency, because a pipeline that copies every finished frame between cards and presents it on a different monitor has an obvious place to hide input lag. OC3D's read, that NVIDIA will keep optimizing the model rather than bless a two-card configuration, is the sensible bet.

DLSS 5 loads on Radeon, and then charges for it

A separate mod, DLSS-NR-on-AMD by danielblnc, gets NVIDIA's neural rendering DLL running on Radeon for the first time. It needs a DirectX 12 game supporting FSR 3 or FSR 4, since the effect is toggled through an FSR profile, plus nvngx_dlssnr.dll in the game directory. Cyberpunk 2077 and GTA V Enhanced are the tested titles, and anti-cheat blocks the DLL outright, so single-player only. The two outlets covering it disagree on reach: KitGuru describes it as RDNA 4 only while Club386 lists RDNA 4 and RDNA 3, worth checking against the project before installing.

The performance is bad enough to be interesting rather than useful. Early builds took an RX 9070 XT in Cyberpunk 2077 at 1080p from above 80 FPS to 11 or 12, and the gains reported since, roughly 12 percent by one account and 26 percent by the other, do not change the category of result. It runs at all because of FP8: DLSS 5's model uses it and RDNA 4 supports it natively, the same property that put the runtime on RTX 40 before NVIDIA committed to supporting it there. danielblnc calls it a ground-up reimplementation of the neural runtime rather than a shim, claims output quality close to the official version, and says some effects still render incorrectly.

Ada gets 4x frame generation back

MFGAdaUnlock patches the memory restrictions that hold RTX 40 cards to 2x DLSS Frame Generation, letting NVIDIA's own DLSS-G code run 3x and 4x on Ada. NVIDIA's 310.9 DLSS-G provider briefly broke it by renaming the temporal kernel and changing descriptor layouts; version 0.4 of the unlock restored Ada temporal-position correction and fixed frame-pacing and duplicate frames at the higher multipliers, per KitGuru. None of it is supported, so latency and stability will not match the tuned Blackwell path, and that 310.9 break is the shape to plan around: an unlock that works today is one provider-side rename away from needing another patch.

Sources