NVIDIA has officially confirmed that it will extend support for its DLSS 5 neural filter to previous-generation graphics cards in the GeForce RTX 40 series. It was previously thought that the heavy-duty generative filtering model of finished images would officially become a Blackwell architecture (RTX 50) exclusive feature. However, after a series of leaks and recent experiments by the modder community, who successfully ran the leaked nvngx_dlssnr.dll library on RTX 40 by replacing CUDA binaries, the “greens” stopped pretending and confirmed that the official implementation will indeed affect the Ada Lovelace generation.
DLSS 5 models aggressively use low-precision FP8 calculations. The fourth-generation tensor core in the Ada Lovelace chip has native FP8 hardware acceleration, which makes the architecture fully compatible with new network inference without the need for deep simulation. In the official release, NVIDIA engineers committed to implementing native execution profiles for Ada Lovelace and adjusting the distribution of model weights to minimize frame time overhead.