Uploaded March 2025 | Updated September 2026, 36 minutes ago
FSR 4 was revealed at AMD's RX 9070 launch event where they unveiled their new RDNA 4 based graphics cards. Transformer models like DLSS 4 have excellent image quality and I wanted to know if FSR 4 would match its quality. I read some recent research papers talking about new unique types of "hybrid" neural network models for image upscaling. I discuss three of those in this video while discussing their implications for FSR 4. Will FSR 4 beat DLSS 4? Probably not. Will FSR 4 be as good as DLSS 4? Nearly as good, but not the same, based on my research.
AMD also showed off a real time path tracing demo called "Toyshop" that utilized neural super sampling and denoiser, a technique AMD wrote about in a paper a few months ago. gpuopen.com/learn/neural_supersampling_and_denoising_for_real-time_path_tracing
AMD seems to be catching up in ray tracing, although reflections in the Toy Shop demo looked very wobbly and like they were "boiling" and I also noticed that the robot didn't appear to be reflected in the ray traced reflections. I think the mirror the robot looks into halfway through the demo is not ray traced and is actually some kind of planar or render to texture technique, because it looks so different than the rest of the path traced reflections. And because of the robot not being in the glossy reflections on the table under the mirror or in the other shiny objects.
0:00 - 0:49 AMD RX 9070 Launch Event
0:49 - 2:14 FSR 4 a Transformer model?
2:14 - 3:49 Official AMD FSR 4 Information
3:49 - 8:14 Recent Research Paper Evidence
8:14 - 10:37 RDNA 4 Ai Acceleration in Compute Unit (Ai Accelerators)
10:37 - 10:54 FSR 4 Overview
10:54 - 12:13 Neural Super Sampling Denoiser
I will be making a follow up video about the neural supersampling and denoising tech AMD talked about, because I had to cut this video short. I spent days/weeks researching for this video and learning a lot about machine learning and RDNA compute units, so if you wanna hear more about that please leave a comment and subscribe to my channel!!
Here's a link to the document I created to collect all the info for this video:
docs.google.com/document/d/12ZRGkgjG4bqRoiVATO0KIMT8SLKupFn1XqetkbLjH2s/edit?usp=sharing
#fsr4 #rx9070xt #dlss4 #rx9070 #radeon #rtx5070
FSR 4 was revealed at AMD's RX 9070 launch event where they unveiled their new RDNA 4 based graphics cards. Transformer models like DLSS 4 have excellent image quality and I wanted to know if FSR 4 would match its quality. I read some recent research papers talking about new unique types of "hybrid" neural network models for image upscaling. I discuss three of those in this video while discussing their implications for FSR 4. Will FSR 4 beat DLSS 4? Probably not. Will FSR 4 be as good as DLSS 4? Nearly as good, but not the same, based on my research.
AMD also showed off a real time path tracing demo called "Toyshop" that utilized neural super sampling and denoiser, a technique AMD wrote about in a paper a few months ago. gpuopen.com/learn/neural_supersampling_and_denoising_for_real-time_path_tracing
AMD seems to be catching up in ray tracing, although reflections in the Toy Shop demo looked very wobbly and like they were "boiling" and I also noticed that the robot didn't appear to be reflected in the ray traced reflections. I think the mirror the robot looks into halfway through the demo is not ray traced and is actually some kind of planar or render to texture technique, because it looks so different than the rest of the path traced reflections. And because of the robot not being in the glossy reflections on the table under the mirror or in the other shiny objects.
0:00 - 0:49 AMD RX 9070 Launch Event
0:49 - 2:14 FSR 4 a Transformer model?
2:14 - 3:49 Official AMD FSR 4 Information
3:49 - 8:14 Recent Research Paper Evidence
8:14 - 10:37 RDNA 4 Ai Acceleration in Compute Unit (Ai Accelerators)
10:37 - 10:54 FSR 4 Overview
10:54 - 12:13 Neural Super Sampling Denoiser
I will be making a follow up video about the neural supersampling and denoising tech AMD talked about, because I had to cut this video short. I spent days/weeks researching for this video and learning a lot about machine learning and RDNA compute units, so if you wanna hear more about that please leave a comment and subscribe to my channel!!
Here's a link to the document I created to collect all the info for this video:
docs.google.com/document/d/12ZRGkgjG4bqRoiVATO0KIMT8SLKupFn1XqetkbLjH2s/edit?usp=sharing
#fsr4 #rx9070xt #dlss4 #rx9070 #radeon #rtx5070
![Doom Eternal (Nekravol) [GTX 1060 6GB][i5 4690k@4.4ghz][Windows]](https://i.ytimg.com/vi/eah_22ci9VU/mqdefault.jpg "Doom Eternal (Nekravol) [GTX 1060 6GB][i5 4690k@4.4ghz][Windows]
Minor spoilers for the Nekravol level. I cant quite get a locked 60 with my setup, even with the settings lowered. So I play this game with a 50fps cap on a mix of low/higher settings. This game looks great regardless so low still looks quite good. With all that, it runs great! There is an added stutter from my recording software, which is not present in gameplay.")
