Basically as per the XeSS for Intel, XMX increases performance and allow more process which increases image quality. This is the same for DLSS and tensor cores. From Intel slides, XeSS on DP4 instructions runs on all GPUs but is both slower and has reduced image quality. One of the big issues with temporal upscaling is a lack of processing power leads to blurring. Also AI is better at removing artifacts and can make guess when there is a lack of samples further improving image quality.
In the Matrix Awakened demo DLSS quality mode appears to outperform TSR in UE5 by 10 fps. So I expect DLSS and XeSS to both outperform TSR and FSR 2 in both quality and performance with a 4k upscale. With DLSS having a fixed cost because the number of tensor cores. I expect the performance to be close at lower resolutions. The issue I would guess with FSR 2 is it wot upscale from 1080p to 4k because the quality wont match. Dynamic resolution scaling is going to be the best bet to keep quality close the nvidia DLSS. you samply increase the resolution or decrease it. Thus you can have better quality and maintain performance. With temporal upscaling there should be little need for sharpening.
In the article we can
see that a 4k result requires a 1440p input for FSR 2 to upscale to 4k in quality mode. Like DLSS this is better than FSR 1. Internal resolution affects the number of rays needed in Ray Tracing games thus rendering at the lowest resolution posible is very important before upscaling. I can see that FSR 2 i about 5 fps behind, given that there is a performnce gap in RT games as well. This means AMD is cannot close the gap with NVidia.
Image quality is good in still shots, need to see raw video to see image quality while the image is moving. Thats when most of the artifacts happen. FSR 2 looks a little more blurred compared to DLSS but close. DLSS 2 does looks closer to native. Stills wont tell you much. Would not use FSR 1 but would use FSR 2. FSR 2 is the right balance, were FSR 1 was not.
