A modder has ported NVIDIA's DLSS 5 neural renderer from using FP8 exclusively to using a hybrid FP8/NVFP4 approach. Reported by SwurvGaming, performance from this DLSS 5 mod only increased performance by 1% to 2% using this approach, despite the raw potential NVFP4 has over FP8 in AI compute.
The modder of this port, wilsjo2, gives details on how this hybrid implementation works. In the patch notes for their OptiScaler DLSS neural rendering mod version 0.7.1, the model selector will attempt an NVFP4 hybrid where possible, with unsupported shapes reportedly staying processed in FP8 when this hybrid mode is active. In version 0.7.2, the modder changed their NVFP4 hybrid model to be the recommended option for Blackwell GPUs, as well as providing some benchmark results. In a 120-second capture of Baldur's Gate 3, the hybrid model in version 0.7.2 achieved 55.16 FPS, and the FP8-only mode achieved 54.57 FPS.
There's not much reason currently to use this hybrid approach with the NVFP4 model, due to its poor performance uplift over DLSS 5's default FP8 implementation. But it shows that modding DLSS 5 to NVFP4 is possible, and we could see better results as optimizations are made to the DLSS 5 model in the future. Porting DLSS 5 from FP8 to NVFP4 is not as simple as switching what instructions the model uses; the model needs to be retrained to take advantage of NVFP4's inherent performance advantages over FP8.
- Read more: DLSS 5 now works on RDNA 4 GPUs through a mod, RX 9070 XT runs DLSS 5 in Cyberpunk 2077 at abysmal frame rates
- Read more: YouTuber shows Vega GPU running FSR 4 with experimental build of OptiScaler and leaked Proton version of FSR 4.1.1
- Read more: DLSS 5 is officially coming to RTX 40 series GPUs after optimizations are fine tuned for RTX 50 series
NVIDIA has already hinted that DLSS 5 will be ported to NVFP4 in the future for Blackwell GPUs. TechPowerUp reportedly asked NVIDIA in its DLSS 5 technical preview if DLSS 5 was NVFP4-quantized, to which NVIDIA answered "not yet". NVFP4 is a new Nvidia-specific 4-bit floating-point format that requires significantly less memory and compute compared to FP8 and BF16 on Blackwell GPU hardware, and can be used for both AI training and inference.
Frequently Asked Questions
Open a question for an answer from TweakTown's coverage of this news, or ask your own below.
Which RTX 50 series cards can use the DLSS 5 NVFP4 hybrid mode in OptiScaler 0.7.2?
Will NVIDIA's promised fall DLSS 5 models include NVFP4 support for Blackwell GPUs?
What changed in OptiScaler version 0.7.1 versus 0.7.2 for the DLSS 5 NVFP4 hybrid model?
Have a question about this content not listed here? Ask below and TweakBot will answer it.
With how demanding DLSS 5 is on even NVIDIA's flagship RTX 5090, NVIDIA needs all the optimizations it can get to run DLSS 5 in a more performant manner on its GPUs. Modders have already uncovered some optimizations to run DLSS 5 faster, which include running the neural renderer before DLSS upscaling in the 3D pipeline instead of after. But even with this performance hack, DLSS 5 is much heavier to run than preceding versions of DLSS. NVIDIA has already promised that performance improvements are on the way, with new DLSS 5 models arriving this fall, so we could see NVFP4 versions of DLSS 5 introduced then.
