When Nvidia unveiled GeForce RTX 5000 graphics in January, various new features were presented (though not all of them are exclusive to these new GPUs), most notably DLSS 4 able to generate more interpolated frames. We’ve devoted a separate article to Blackwell’s features, but now that the GPUs have started selling (albeit in limited quantities), we see that that are some additional new features that have flown under-the-radar before.
We’ve already covered Smooth Motion technology, which extends frame generation (though not the multi-frame mode) to games that don’t support this feature or DLSS at all. Smooth Motion runs purely within the driver level without game integration, meaning it can be enabled universally. However, this comes at the cost of lower quality, as it lacks access to in-game data about motion vectors, GUI elements, and other metadata. This technology is akin to AMD Fluid Motion Frames, which is based on FSR 3 frame generation. Radeon GPUs were the first to market with introducing this kind of universal driver-level implementation.
Alongside Smooth Motion, the driver update that was released with the launch of the GeForce RTX 5000 series introduces another new feature—an improved AI-powered video upscaling for the feature called RTX Video Super Resolution, which originally launched in 2023. The latest drivers bring a new version of this feature, though Nvidia has not assigned it a specific version number, simply referring to it as the “January 2025 update.”
This year’s RTX Video Super Resolution update follows a similar pattern to DLSS 4 in that it incorporates a newly trained neural network for video upscaling and post-processing. With DLSS 4, Nvidia introduced a new transformer-based neural network, which should enhance the quality of generated frames. However, there’s no mention of a shift to transformer networks for this updated RTX Video Super Resolution, suggesting that it does not use the same model as the gaming upscaling technology does.
Better Performance and HDR Support
The announcement of this update also does not highlight any improvements in image quality like those that are advertised for DLSS 4, so users shouldn’t expect significant (if any) visual enhancements compared to prior version. Instead, the January update focuses on boosting performance, or rather efficiency. The new AI model thus seems to be focusing on this metric specifically. Nvidia states that the new neural network reduces computational demands by about 30%, lowering GPU load by the same margin.
The second major change is that upscaling now officially supports HDR video. However, this might be a capability that was not sorely missed before, as HDR content is typically already in high resolution, and heavy-duty upscaling with post-processing is generally more beneficial for lower-quality, low-resolution videos. When dealing with high-quality, already sharp and compression-free video with a high level of detail, aggressive post-processing might do more harm than good.
Additionally, Nvidia has introduced an on-screen indicator for when upscaling is active, along with a separate indicator for the automatic HDR conversion feature. This can be useful when testing in an application to confirm whether the function is truly enabled. And notably there is also a new option to automatically adjust the quality (intensity) of scaling.
Automatic quality and priority
RTX Video Super Resolution allows users to choose between multiple levels of upscaling intensity. There is also an “auto” setting, which dynamically adjusts the upscaling quality to match the GPU’s capabilities – up to a certain point of course, it is likely that even he highest mode the technology offers won’t really exhaust the performance of GPUs, at least with higher-end ones.
A new addition to this automatic mode is the “priority” setting (low/medium/high). This lets users define how the GPU should allocate resources when video upscaling runs alongside other tasks. A higher priority ensures the best possible upscaling quality for RTX Video Super Resolution is always used, while the low priority setting should ensure that a high quality is used only when you are not using the graphics card otherwise, but should the graphics card be under other load, the quality of upscaling is reduced, to avoid slowing down say a game or a productivity tasks, for example.
These improvements are now available in the latest Nvidia App release (11.0.2.312) and are applicable to all GPUs that supported the previous version. The feature isn’t exclusive to the new generation of GPUs, but since it relies on Tensor cores, it’s only compatible with GeForce RTX 2000 series and newer cards.
Source: Nvidia
English translation and edit by Jozef Dudáš
⠀
