DLSS 5 Mixed-Precision Mod Shows Only 1-2% Performance Gain on RTX 50 GPUs
Modders testing FP8 and NVFP4 mixed-precision optimizations on leaked DLSS 5 neural rendering models for RTX 50-series GPUs observed only a 1% to 2% reduction in neural rendering execution time. Implemented in versions 0.7.1 and 0.7.2 of the OptiScaler project, the modification attempts to leverage Blackwell Tensor Cores by converting part of the workload from FP8 to 4-bit NVFP4. Despite hardware-level native support for NVFP4 in NVIDIA's Blackwell architecture, simple precision quantization of existing DLSS 5 models fails to deliver major performance leaps. This suggests that achieving meaningful speedups in real-time neural rendering will require deeper architectural changes, model quantization tuning, and driver-level optimizations. In a 120-second benchmark of Baldur's Gate 3, frame rates increased marginally from 54.57 FPS with FP8 to 55.16 FPS with the NVFP4 mixed-precision scheme, a gain of roughly 1.08%. Additionally, model shapes not supported by NVFP4 automatically fell back to FP8 execution, preventing the mixed-precision pipeline from covering the full calculation process.
## BACKGROUND
Neural rendering integrates deep learning models directly into the graphics rendering pipeline to enhance image quality and performance. NVIDIA's Blackwell GPU architecture introduced 5th-generation Tensor Cores with native support for NVFP4, a 4-bit floating-point format designed to reduce memory bandwidth and inference overhead compared to FP8.