Bummer, AMD has over 1 year time to move RDNA 4 to Zen 6. I guess it takes more time? I don't think AMD is doing it on purpose, although even Zen 5 still only has RDNA 2 (thanks AMD for giving me another reason to not upgrade from my Zen 4). I understand that in a desktop CPU, its iGPU matters much, much less than for mobile, as a desktop gonna have a dedicated GPU anyway.
According to:
computerbase.de/artikel/grafikkarten/amd-radeon-rx-9070-xt-rx-9070-test.91578/seite-9#abschnitt_energieeffizienz_in_fps_pro_watt
the power efficiency improvement of 7900XTX to 9070XT is only 7%, but 70% if using a FPS limit!
Indeed, supposedly FSR4 uses FP8, which is only in RDNA 4.
Someone said:
QuoteThe current FSR4 is a hybrid between a CNN and a Transformer model from a whitepaper review with sources here, and in theory a FSR4-lite can exist by not utilizing the Transformer at all. Essentially near-parity to DLSS3 without the ability for the FSR-Lite to use the Transformer layer, but only if there's additional work done to resolve the sharpening issues.
Transformer model needs FP8 that RDNA 3.X lacks, but the dedicated existing INT8 might be utilized better with better model training in the future. Or even a Transformer model that somehow uses those same INT8 fixed-fuction hardware to run.
So maybe INT8 in RDNA 2 and RDNA 3 can be utilized for FSR 4 (the performance and/or quality could be different tho).