Multi-frame generation is only available for RTX 5000. Nvidia’s DLSS chief is now opening up the possibility that regular frame generation could also run on older GPUs in the future.
In the course of CES 2025, Nvidia not only presented the new GPU generation around the RTX 5090, but also introduced the new version of its in-house AI upscaling with DLSS 4.
The catch: the feature of DLSS 4 called “Multi Frame Generation”, which generates up to three frames per intermediate image calculation and thus promises a significantly higher frame rate per second, will be reserved exclusively for the new graphics cards.
Nvidia has already done something similar with the first version of frame generation, which came onto the market with DLSS 3. This technology could only be used on the RTX 4000, but that could soon change.
Nvidia leaves room for speculation
Nvidia’s Vice President of Applied Deep Learning Research, Bryan Catanzaro, recently answered questions from Digital Foundry in an interview.
In addition to an interesting deep dive into the new DLSS 4, the question was also raised as to whether frame generation on GPU generations like the RTX 3000 would be technically possible. Catanzaro commented as follows:
I think it’s mainly a question of optimization, engineering, and ultimately the ultimate user experience. We’re introducing the best multi-frame generation technology with the RTX 50 series, and we’ll see what we can get out of older hardware in the future.
With that, Catanzaro opens the door a crack for retroactive frame generation support for RTX 3000 (and maybe even older GPUs), even if the Nvidia executive couldn’t be drawn to give a definite answer.
The stumbling block for these mind games was a previous question about the fundamental change that Nvidia is making to its upscaling technology in the course of DLSS 4.
- Because instead of relying on the so-called “Optical Flow” hardware accelerator (OFA), DLSS 4 uses the tensor cores of the GPUs and is also supposed to use less video memory.
- So far, Nvidia has argued that older architectures such as Ampere and Turing do not have sufficient OFA performance to support frame generation technology.
In theory, the Catanzaro statement opens up another scenario: Multi Frame Generation should – provided that only a sufficiently high tensor performance is actually required – also be able to run on high-end GPUs of the RTX 4000 series.
However, it is doubtful that such a step will be taken, at least for the next few months, because from a market strategy point of view, it would be shooting itself in the foot when the RTX 5000 generation is released.
So the most likely scenario would be one in which the change to tensor cores as the basis for upscaling would equip older graphics cards up to the RTX 30 with the “first” version of Frame Generation.
A “feasibility study” for this has been provided not least by AMD, which has also been fielding an FG variant since the introduction of FSR 3 – and is unlocking it open-source for all FSR-3-compatible GPUs.
