Odpověď na názor

pro j. olsana:
"[tensor core] makes it possible for us to feed in a lower resolution image, scale it up in real time and then reconstruct missing details. And not just sharpen the result, but improve the result. Actually make the details cleaner in some cases. And it does this because .. sort of the best explanation it's a set of AI algorithms that have been trained to know what things should look like. It was trained on very high resolution images, like 64K images, very high resolution photographs. Taught how light should behave, how detailed it should look. And all this training was done for a very long time to teach the code really so it could tech itself how images should appear perceptually to the human eye, and then fill in the blanks. So it's not just upscaling it and sharpening it. It's doing a lot more than that."

"The artificial intelligence is inserting and creating enhancing details."

"Q: Can you use this even without raytracing enabled?"
"A: Oh yes, the scene doesn't need to be raytraced at all. It's for any kind of image. Raster, raytraced, any mix inbetween. The only requirement from a strict requirement side is that it requires RTX hardware to run. There's not much we can do about that because in order to reconstruct that image really quickly, you need Tensor Cores. You need some type of neural net hardware basically. Something that's really fast and really powerful, capable of doing billions of operations per second."
https://youtu.be/a9KJo9llIpI?t=1583

takze tady to je black on white, primo od cloveka z nvidie: "So it's not just upscaling it and sharpening it. It's doing a lot more than that." "[dlss is] a set of AI algorithms that have been trained to know what things should look like .. taught how light should behave, how detailed it should look .. and then fill in the blanks".

tak co olsane, porad budes tvrdit, ze dlss je upscaling? ocividne funguje presne jak jsem tehdy rikal. uci se detaily obrazu, ktery pak doplnuje :D

btw ta druha cast je zajimava taky - i kdyby nvidia chtela zpristupnit dlss na amd hardwaru tak nemuze. opet jak jsem psal uz davno, pocitat tak velkou dnn na klasickych shaderech by sebralo tolik vykonu, ze by pak nezbylo nic na klasickou rasterizaci (prinost pro vykon by byl nulovej, spis pravdepodobne i zapornej). a to je tak duvod, proc fsr neni zalozeny na nejaky komplexni ai, tzn. ne ze by amd melo nejakou tajnou formuli jako delat to, co nvidia, jednodusejc. oni proste nemuzou, protoze na to nemaji hardware. zaroven to vytvari dilema pro amd, protoze se muze stat, az nekdy amd uvede obdobu tensor cores, ze to nvidia zpristupni i tam a fsr zabije (protoze maji roky naskok v trenovani jejich nn)