AMD has confirmed plans to open up its acclaimed FSR 4 upscaling tech so it runs on competitors’ GPUs, including those from Nvidia. At a roundtable about FSR Redstone at CES 2026, we asked Josh Hort, AMD’s senior director and head of ISV enabling for computing and graphics, whether the future of FSR upscaling could be GPU-agnostic and run on Nvidia Tensor cores. ‘Absolutely,’ he confirmed. ‘It could run on anybody who has the performance and throughput to run the model.’
‘Andrej [Zdravkovic] commented that we do have plans to open at least the upscaling portion [of FSR Redstone] at some point,’ Hort added. This referenced an earlier interview about Redstone between PC World and AMD’s senior vice president of GPU technologies and engineering software, Andrej Zdravkovic, which you can see in the video below. ‘I can’t comment on when exactly,’ says Hort, but there’s very clearly an appetite to get FSR Redstone upscaling working beyond the AMD stable.
AMD has often opened up its tech in the past, including its FreeSync technology, as well as earlier upscaling systems. ‘We have open-sourced previous versions of FSR,’ points out Hort, ‘and actually, because we’ve done that, Microsoft selected it as the foundation for their Auto SR [Automatic Super Resolution] technology in Windows.’
Hort says that ‘Andrej’s philosophy is the more open, the better.’ Are we just talking about upscaling here, though, or could AMD also open up its other FSR Redstone tech, such as ray regeneration and FSR frame gen?
‘I think you’ll see in the future it’ll probably be on a feature-by-feature basis as to what’s open and when,’ says Hort. There’s an elephant in the room here, though, which is that while there’s talk of opening up FSR Redstone upscaling to work on competitors’ GPUs, it still doesn’t officially work on AMD’s older RDNA 3 cards.
Unlike previous versions of FSR, the upscaling part of Redstone, previously known as FSR 4, uses hardware-based machine learning. Specifically, it requires AI hardware with support for 8-bit floating point (FP8) instructions. These are supported by the AI Accelerators in AMD’s latest RDNA 4 GPUs, and Nvidia’s Tensor cores from Ada (RTX 4000 series) onwards. However, they’re not supported by AMD’s earlier RDNA 3 GPUs, or Nvidia’s Ampere and Turing GPUs.
That hasn’t stopped Nvidia from making its new DLSS Transformer model run on its Ampere and Turing GPUs, though, despite it also using FP8. You still get the improved image quality of DLSS Transformer on these older GPUs, but there’s a notable performance hit. Are there really no plans for AMD to allow FSR 4 to officially run on older GPUs, just with slower performance? Plenty of modders have shown you can get it working after the accidental FSR source code leak in 2025.
‘It’s one thing to take leaked source code and run it on the GPU, but it’s another thing to make it a real product,’ says Hort. ‘There’s a lot of QA that gets involved, a lot of testing, other optimisations, right? That code – it was somebody’s sandbox that should not have been published. We’re always evaluating our roadmap and what we can support, given our resources, but…we specifically made these models for FP8, and assuming that you get the ML [machine learning] throughput of RDNA 4 in that architecture.’
That’s not a no, but it does look as though it’s not a priority for AMD at the moment. We’re looking forward to seeing how well FSR Redstone upscaling works on other manufacturers’ hardware, though.
