So here's Nvidia's "final say" with DLSS 4.5:
- Even though it's 20 Series and 30 Series compatible, there's a massive performance hit because they don't support FP8 precision
- In my mind, there's no point in even considering using DLSS 4.5 for these older cards
- Model M (as well as L) benefits over past models in any mode, but the performance hit is significant enough, even on 40 and 50 Series cards, that Nvidia still recommends DLSS 4 for DLAA/DLSS Quality/DLSS Balanced for the best quality/performance ratio
- So Model M is indeed "better", but the performance hit is too large for Nvidia to recommended it with the higher quality scaling options
- Ray Reconstruction has not been updated to DLSS 4.5, so if you turn it on in any game, you will be held back to DLSS 4.0/1st gen transformer architecture regardless of your other settings
- This is just dumb. Why release DLSS 4.5 if it's not ready across all of the RTX featuresets?
Long story short is if you have the head room, Model M does appear to be the way to go for increased image quality, motion clarity, and temporal stability. But the performance hit is large enough that Nvidia doesn't feel comfortable advertising it as the "set it and forget it" model.
Honestly it's a cluster fuck and I have no idea why they even rolled this out if there's so many stipulations. As a 5090 owner I'm sticking with forcing Model M though.