By using this site, you agree to our Privacy Policy and our Terms of Use. Close
haxxiy 19 hours ago
Edit Thread
Jizz_Beard_thePirate said:

According to techpowerup, FSR4 needs an AI acceleration throughput of 779 AI TOPS. RX 9070 has around 1200 TOPS, RX 9070 XT has around 1500 TOPS. PS5 Pro has around 300 TOPs for reference but supposedly FSR4 is using a different implementation than Sony and Nvidia to achieve it's Ai upscaling. Frame Generation is the same as FSR 3 so it's not Ai generated but honestly, the Ai generation part of Nvidia's Frame Gen hasn't impressed me anyway. At the end of the day though, the proof will be in the pudding.

I suppose that's just them referring to the overall performance of the 9000 series cards because I'm almost positive FSR4 isn't running on sparse INT4 instructions but FP8 instead, which would be ~290 and ~390 TOPS respectively for the 9070 cards.

That is more comparable in scope to the PS5 Pro (which Cerny called "8-bit TOPS" but I reckon it means FP8/BF8) and the computational costs of DLSS 3.5 CNN at higher resolutions.

So, theoretically, they could run it on the PS5 Pro. I'm just not sure/remember exactly how it does matrix calculations (if it's more like RDNA3 or RDNA4) since one takes away from raster ops and the other seemingly does not.