Modders Unlock DLSS 5 and Multi-Frame Generation for Older Nvidia GPUs

Breaking Nvidia's Software Locks

Over recent weeks, the modding community has achieved what Nvidia claimed was impossible: running DLSS 5 Neural Rendering and Multi-Frame Generation (MFG) on older GPU architectures. Two modders, Dashdogy and Uncle Burrito, have broken down the software-based locks that restricted these features to Nvidia's latest Blackwell-based RTX 5000 series GPUs. The foundational mods now serve as the baseline for a wave of forks and integrations that have extended support all the way back to 2018's Turing architecture and, in DLSS 5's case, to AMD hardware.

Unlocking Multi-Frame Generation

Dashdogy, an Australian student studying Physics, Mathematics, and Space Science, spent a week during mid-semester break reverse-engineering MFG after the feature was never released for Ada Lovelace hardware despite being discussed for RTX 4000 series cards. The key breakthrough came from targeting Streamline, the universal translator between games and upscaling technologies, and Nvidia's proprietary Neural Graphics Framework (NGX). By spoofing the memory calls that define how many frames the GPU can generate, Dashdogy tricked Ada GPUs into believing they could generate five frames instead of one, bypassing the black screen error that had blocked earlier modders.

An initial release packaged as a Cyberpunk 2077 mod revealed frame pacing problems at multipliers beyond 2X. Dashdogy used a cloud instance with an RTX 6000 Pro Blackwell GPU and analyzed kernel dumps with the help of an AI model to identify the root cause: the 40-series kernel was hardcoded to output midpoint scaling for 2X only, compressing generated frames toward the middle of the frame interval. The fix spread the frames properly to match the Blackwell reference, and the mod now works as a loadable ASI file supporting up to 6X generation, with forks already integrating the logic into ReShade.

Porting DLSS 5 Neural Rendering

Uncle Burrito, an experienced graphics enthusiast who previously implemented Path Tracing into Need For Speed Underground 2 using RTX Remix, set to work when DLSS 5's nvngx_dlssnr.dll first appeared online. Although written for Blackwell, the Ada architecture shares FP8 instruction support needed for the model weights. Of the 15 CUDA binaries in the DLL, only four were incompatible with Ada. Using the reverse engineering tool Ghidra and Claude Code, Uncle Burrito recompiled the problematic binaries, though an initial attempt failed because the AI model selected the wrong CUDA SDK version, producing a black screen. After recompiling correctly, the patched DLL ran on RTX 4000 series GPUs. Nvidia later clarified that Neural Rendering would indeed be supported by RTX 4000 series cards, lending credence to the idea that Ada support was unfinished rather than technically impossible.

The Broader Impact

Because motion vectors and depth buffers are standardized across modern games, the patched DLL functions as a drop-in replacement for the officially distributed one. Installation is now possible through click-and-install tools across hundreds of games. The irony is that both unlocks relied on AI models, the same technology driving Nvidia's business, to dismantle its own software restrictions. With RTX 5090 GPUs now approaching $9,000, the mods offer owners of older cards a way to extend the life of their hardware. For those unafraid to modify a few files within a game, these features are now available, albeit unofficially.

Sources: Tomshardware