AMD FSR antialiasing discussion

Man i'm such a dummy. I didn't even think about using other forms of visual info, ie feature maps, as an input into your final NN.
You could even do a simple edge detection filter and then turn that into a vector based descriptor for a resolution independent description of edges.
I guess thats how you get nice clean edges on your upscaled image.

I wonder what other kinds of feature maps they could/would be using as inputs into the upscale process.
It's a real pity they have any access to an Engine level depth map, which would be super useful for this stuff
I guess they could generate it, but no where near as accurate.

 

Thanks, I am aware of macroses. Slower speed dp4a would have been noticed immediately and caused a lot of questions from CUDA community, but I don't see any.

 

Having looked into this further, this does seems like a documentation issue on Nvidia's part. I mean, who in the right mind would not expect the hardware vendor's own Tuning Guide to be the source of truth of their own hardware?

 

The Tuning guide doesn’t explicitly say anything about other GP10x chips. So it can be neither right or wrong

 

The authors of said tuning guides should cross-check with the executives of product segmentation. Not all features are accessible in all SKUs. That being said, I was unaware of differences in SMx functionality between x02, x04 and x06.

 

Unreal Engine 5 demo is using AMD Temporal super resolution

https://community.amd.com/t5/blogs/...s-in-unreal-engine-5-early-access/ba-p/473407

 

Seems like it's using it's own temporal supersampling though - which is already present in UE4 but is expanded in UE5? The wording isn't clear. They would probably call it FidelityFX SR if it was that.

 

Its a UE5 engine feature, hence it running on anything.

 

Well, "wording" is pretty similar... at least in UE5 release notes - Super Resulution, Fidelity etc.

 

FSR is expected to run on anything too.

But it does invoke an interesting question: would UE5 (and other "big" engines) even integrate FSR if their own solution will end up being about 90% similar? Would there be any benefit to this?

 

This is not AMD but Epic solution.

 

Sorry wrong thread

 

They're teasing FSR from Computex ( AMD Computex thread @ https://forum.beyond3d.com/threads/computex-2021-amd.62408/ )

 

Apparently it's a spatial-only filter..

 

To use "motion vectors" is only viable with specialized hardware like tensor? Or normal compute power is enough?

 

TAA has been using motion vectors since its inception almost 10 years ago.

 

RDNA3 and up will see improvements im sure.

 

I remain skeptical until I see some independent video comparisons.