I definitely see that as the safe bet, though, just looking at the progression of things here for AMD and MS:
- FSR has traditionally hit a very wide range of hardware, narrowing this down to RDNA 4 would be very specific, and if few people purchase RDNA4, then all this investment to train a model is worthless.
- AMD made an announcement to cede the high end market
- All their mobile devices and above are approximately DX12U and above now
- MS releases DirectSR - which effectively abstracts super resolution code so that developers can code once to support a variety of upscaling algorithms. This would be extremely handy for ML upscaling and fallback requirements for lower end systems, it could allow for AMD to support multiple variants of ML algos.
- MS owns CoD Black Ops 6
- AMD announces ML upscaling for Black Ops 6, it is already optimized today for FSR3.1
- The Series consoles were delayed for RDNA2 -- this may have been a for a dp4a requirement, I would be surprised that FSR being the default solution for DirectSR, the Xbox consoles would be excluded from running the ML variant.
- We know XeSS can run on DP4A and DLSS1.9 was entirely compute based without DP4A
Lastly from this article:
AMD's next-gen FidelityFX Super Resolution technology known as FSR 4 will be leveraging AI in a grander scale to improve visual fidelity.
wccftech.com
Sounds like AMD is not referring to RDNA4 here, and are intending to going wide.
IMO, the only reason I don't think we've seen ML algorithms on all GPUs is because Nvidia keeps it locked to tensor cores now, they are the premium market leader, they will force their users to upgrade to the latest and greatest. XeSS is not quite open source, and AMD had not completed making their own, these are costly models to train. Otherwise I think we could have had ML upscaling available on all sorts of hardware. I've never seen DLSS as a tensor core requirement. We also don't have any benchmarks on how long these networks take to run on hardware, it's not something we have any real insight into. So most of it, is us just relying, that the IHV needs this hardware to be able to run this software. But my experience in the field, this has not been the case, unless model training is involved. That being said, more tensor power would enable you to run denser networks resulting in high quality in terms of output.
PSSR is likely a customized or earlier release of the AMD FSR4.0 property, and we've seen this type of behaviour before, 5Pro will be launching with some RDNA4 tech before it's announcement.
As you stated, the safe bet is RDNA4 and imo RDNA3 devices will definitely be supporting this. The question is how much farther are they willing to broaden it.