Modders make leaked version of DLSS 5 run on RTX 40 cards with frame drops

A preliminary edition of NVIDIA’s long-awaited graphical artificial intelligence system, which ended up unofficially appearing online, already works on Ada Lovelace generation video cards. Thanks to the efforts of software modification enthusiasts since the end of August 2026, the public was able to see the first glimpses of the manufacturer’s new neural rendering library. However, initial trials demonstrated a severe decline in frame rate, sparking heated debates about how the company will handle the compatibility of its new features on components already on the market.
The confidential file’s path into the hands of the modding community
The central component of this innovation, focused on improving images through neural networks, appeared unexpectedly in the files of a test version of the NBA 2K27 basketball simulator. Technically named nvngx_dlssnr.dll, the digital document quickly captured the interest of independent programmers who scour code in search of hidden innovations.
The group of amateur developers known as RenoDX, famous for creating unofficial solutions for computer games, were the first to integrate the leaked file into the acclaimed action title Control. Soon after this initial success, the same testing structure was successfully applied to more than ten different games, proving the high flexibility of the discovered code.
Reverse engineering allows artificial intelligence to function in the Ada Lovelace generation
In its original state, the graphics enhancement tool was locked to operate exclusively on the upcoming RTX 50 family. This restriction was because specific blocks of programming within the library required CUDA instructions custom-built for the Blackwell architecture, preventing it from running on older systems.
It was the independent programmer nicknamed Uncle Burrito, already renowned for his projects involving the RTX Remix platform, who got around this architectural barrier. By mapping the codes that blocked operation on Ada Lovelace chips, he was able to exchange incompatible instructions with functional equivalents, unlocking the use of cutting-edge neural rendering in current high-performance hardware, such as the coveted GeForce RTX 4090 and RTX 4080.
Drastic drops in fluidity mark the first practical tests of the software
Even with the success in breaking the hardware blockage, the application of this preliminary technology took a huge toll on the performance of the 40 series GPUs. Evaluations conducted on the top of the line model, the GeForce RTX 4090, revealed a computational bottleneck that reduced the fluidity of the images in an alarming way.
In one of the graphics stress scenarios, the frame count plummeted from an average of 135 to just 82 as soon as the new neural system was activated. This reduction of approximately 39% in visual processing capacity raised a warning signal among consumers who invested heavily in the brand’s current generation of video cards.
- Evaluation on GeForce RTX 4090: abrupt reduction from 135 to 82 frames per second (39% loss of fluidity).
- Previous tests using the GeForce RTX 5070 in the Cyberpunk 2077 game: decrease from 138 to 68 frames per second.
- Stress tests with the GeForce RTX 5070 Ti in the Control title: sacrifice of practically half of the original performance.
It is essential to note that this unstable behavior is not exclusive to adapted hardware, as similar losses occurred in leaked tests of the RTX 50 line itself, including entry-level components such as the RTX 5060. Although the manufacturer has only made official controlled displays running on a dual RTX 5090 system, independent experiments make it clear that the base technology covers the entire family of Blackwell chips, even if it still lacks refinement.
What to expect from the final polishing before the official arrival on the computer market
The encouragement for technology enthusiasts is that the current code represents just a draft of what will be delivered to consumers in the future. The processing bottlenecks recorded in these leaks are far from reflecting the final product, as the company’s engineers still have months of work ahead to smooth out the edges and optimize resource consumption.
The semiconductor giant’s schedule points to a commercial debut of the new graphical artificial intelligence tool in the last quarter of the year. Until the date arrives, the version intercepted and modified by the RenoDX group works as an open laboratory for programmers to study in depth the mechanics of neural rendering. The true visual and performance impact on the current GeForce RTX 40 will only be known when the developer releases properly calibrated official drivers to the general public.






