Modern SLI? Running DLSS 5 on Two RTX 5060 Ti Cards

How expensive is DLSS 5 neural rendering on the GPU? A ReShade plugin on GitHub called MGPU Bridge has put a number on it: with the game running on the first card and DLSS 5 neural rendering fully offloaded to the second, frame rate jumps from 59 back up to 107.

💡 What You Will Learn

How expensive is DLSS 5 neural rendering on the GPU? A ReShade plugin on GitHub called MGPU Bridge has put a number on it: with the game running on the first card and DLSS 5 neural rendering fully off

How the two cards split work: the setup uses two RTX 5060 Ti 16GB cards. The primary card renders the game; the rendered frames are passed via the ReShade plugin to the secondary card, which runs DLSS 5 neural rendering and outputs the result to a second monitor connected to the secondary card. This has nothing to do with old SLI — SLI splits one frame's work across two cards; here every frame's rendering work is unchanged and uncopied, only the neural rendering step is moved to a different device. Why it works in principle: neural rendering happens after the frame is finished — it takes one finished image in and emits a finished image out, so it can naturally run on another device. Measured numbers (Dawnwalker, 1080p, three columns: NR off / NR on primary / NR on secondary; note that 'NR off' still has NNR's supersampling stage enabled): DLAA 67–70 / 44 / 67–70 FPS; Quality 98–99 / 54–55 / 91 FPS; Performance 127–131 / 59 / 106–107 FPS; Ultra Performance 172 / 69–71 / 157 FPS. Translated into cost: NR packed onto the primary card loses 36% to 59% of frame rate; offloaded to the secondary card, the loss is 0% to 17%. In DLAA mode the offload has zero loss — frame rate matches 'NR off' exactly. The primary card's temperature also dropped by 21°C because the load is split across two coolers. Key finding (the slope): from DLAA up to Ultra Performance, this machine should theoretically deliver a 151% frame rate uplift. With NR crammed onto the primary card, only 39% of that uplift is actually realized; offloaded to the secondary card, 86% is realized. Why: neural post-processing always runs at the output resolution, so the lower the rendering resolution, the bigger NR's share of each frame — the savings supersampling gives back are eaten, bite by bite, by neural rendering. Moving NR off the primary card makes supersampling a real bargain again. A neat detail: during testing the primary card was deliberately plugged into a chipset-lane PCIe 3.0 x2 slot — the narrowest data path on the board — and every frame still had to traverse it to reach the secondary card. The setup still works. Realistic caveats (from the author's README): both cards must be RTX 50 series (RTX 40 is theoretically possible but untested and unsupported); two monitors, one connected to each card (headless secondary loses 33% throughput); only DirectX 12 games are supported right now; frame generation is untested; color processing isn't implemented yet; full testing covers only three games, measured in minutes of runtime. The author's exact words: this is research code, not a product, and not a reason to buy a second GPU.