Sony beat Microsoft in Machine Learning for Consoles by doing
something, which is more than the
nothing that Microsoft did on their end.
Microsoft announced ML capabilities for upscaling back in 2020 and made a big fuss about having DP4a capabilities (8x throughput on UINT4, ~96 TOPs on SeriesX) for that end, after which they failed to release anything at all. Not even a demo or a trailer, and the Xbox team
loves their tech demos and trailers (just look at
Coalition's show-off in early 2021 around variable rate shading for a game that still doesn't even have a release year).
Last month they announced DirectSR which is basically a tool to implement FSR, XeSS and DLSS in DX12 games, once again showing they have nothing of their own.
It could be that they saw open source initiatives like FSR2 and Intel's XeSS DP4a approach and decided their own efforts weren't worth pursuing, but in the end it's just another episode of Microsoft announcing big plans about something and then forgetting they ever mentioned it. Like the Xbox One using cloud processing or the One X having big plans for VR.
Unlike the DF team's thoughts, I don't think this has anything to do with performance either. They seem to be convinced that DLSS takes a massive toll on Nvidia's tensor cores
but it really doesn't. It looks like the Series X could do run DLSS2 pretty fine (
and Nvidia does have a compute shader version of it running on the Quadro T series, but ssshhh).
Pretty sure they're talking about Microsoft's announcements back in 2020 regarding ML-based upscaling that was going to make it so much faster than the PS5 (according to
insiders).