I'm not following the threads, so I have no idea if this was already discussed here.
More than a year ago Microsoft discussed their solution for Image Reconstruction Upscaling using DirectML.
[youtube]
www.youtube.com/watch?time_continue=2251&v=QjQm_wNrvVw[/youtube]
Starts at minute 24.
It's interesting that while doing their research Microsoft working with Nvidia, not with AMD.
Nvidia gave them their "secret sauce", but it's not relevant anymore because by the time it was the original version of DLSS, that gargabe. Since than Nvidia learned with AMD's RIS and changed everything to release DLSS 2.0.
Anyway, you can see that it's just a question of software. Just like you don't need dedicated hardware to do raytracing the additional hardware for ML just speeds the work. Work that can be done with low precision math. There Microsoft explains that their solution uses FP16, and I guess it can be done with INT8 and INT4. Guess what, RDNA1 supports double rate FP16 and quad INT8 and octa INT4. Microsoft talked about customization to improve this and we suppose this means this work can be done across all the CUs, and not limited like it is on RDNA1.
So, Microsoft is working on this for a long time, and because they had to work together with AMD it's safe to assume that AMD's solution will be very similar and use the same hardware optimizations as the SeX.
We can also expect this coming to RDNA1, and because it'll be open source Sony will also probably take advantage of this work if they didn't do their own work to offer the same on the PS5.
All consoles must receive this feature via firmware update next year.
As expected this isn't and Nvidia advantage anymore, it's a new industry standard.