DLSS 5 Modders Run Hybrid FP8/NVFP4 on RTX 50 GPUs, but Performance Improves by Just 1–2%

Modders are testing a hybrid FP8/NVFP4 version of DLSS 5’s Neural Rendering model on GeForce RTX 50 Series GPUs. The experiment aims to reduce the technology’s performance cost.

The implementation was introduced in version 0.7.1 of the OptiScaler-DLSSNR-PreSR-Multipass project, according to a post highlighted by X user SwurvGaming. The project’s release notes report that the hybrid model improves complete DLSS 5 performance by only 1–2%.

The approach uses both FP8 and NVFP4. Blackwell’s 5th-generation Tensor Cores natively accelerate NVFP4, making the format a potential way to improve inference performance. The reported result, however, shows only a small gain from moving parts of DLSS 5 away from FP8.