Intel XeSS anti-aliasing discussion

[Frenetic Pony]

The framerate at which interpolation is being used is rarely enough for VR without an extra frame of latency, with the extra frame it's a lost cause.

You can decouple view movement from the main framerate (with a bit of bodging for the stereo) but any interaction other than view movement will still have the latency.

A drop in TAA replacement is too brain dead, renderers have to get a lot smarter for good sample reuse for VR.

Of course sticking a 720p display right up to your eye is going to reveal some TAA artifacts. I'd love to try an Apple Vision Pro eventually, it's what, 90 degree fov over 3660 or something? 40 pixels per degree of arc is enough for TAA to start being fine.

 

[MfA]

Sorry, I meant to say that sample reuse for framerate enhancement which is implemented as a drop in TAA replacement is too brain dead for VR.

Extrapolation will need to be more tightly married to the renderer.

 

[Cyan]

Wait, Intel just gave DLSS to Everyone! - HUGE XeSS 1.3 Update

 

[Dictator]

Wait, Intel just gave DLSS to Everyone! - HUGE XeSS 1.3 Update

I think you will like today's DF video. 1.3 for sure advanced XMX and DP4a versions considerably.

 

[Cyan]

another superb video by Alex, with great attention to detail, sometimes to the point of being like an intense love for detail.

 

[Frenetic Pony]

another superb video by Alex, with great attention to detail, sometimes to the point of being like an intense love for detail.

Waiting like, maybe a week or two, for FSR 3.1 to have complete video of all the updates that were announced around the same time would've made more sense, what a weird decision.

 

[Reynaldo]

Waiting like, maybe a week or two, for FSR 3.1 to have complete video of all the updates that were announced around the same time would've made more sense, what a weird decision.

Are they not allowed to make a new video when FSR 3.1 drops?

 

[Frenetic Pony]

Are they not allowed to make a new video when FSR 3.1 drops?

Comparing New Xess, to New DLSS, to Old FSR is a weird narrative, "look how much better machine learning TAA is!" when the released, assumedly not machine learning, FSR 3.1 is out at the same time.

Give prospective customers of GPUs a fair idea of what they'd be buying for practical purposes immediately across the board, rather than deliberately have a year old out of date comparison point.

Give curious people a survey of all 3 TAA GPU vendor software packages released at the same time a fair comparison, rather than having an old one to "show off" how much better the other ones are. Isn't the improvement generation on generation more of a relevant comparison than 2 new releases compared to an artificially chosen year+ old release?

 

[Reynaldo]

Comparing New Xess, to New DLSS, to Old FSR is a weird narrative, "look how much better machine learning TAA is!" when the released, assumedly not machine learning, FSR 3.1 is out at the same time.

Give prospective customers of GPUs a fair idea of what they'd be buying for practical purposes immediately across the board, rather than deliberately have a 2 year old out of date comparison point.

They will, once they have the data. What's the big deal? They're a channel that needs constant content to make money. What a weird thing to be complaining about.

 

[DegustatoR]

Waiting like, maybe a week or two, for FSR 3.1 to have complete video of all the updates that were announced around the same time would've made more sense, what a weird decision.

Waiting a week or two for 3.1 and then waiting for how long for a solid number of implementations which would make enough for the proper comparison selection?

Comparing New Xess, to New DLSS, to Old FSR is a weird narrative, "look how much better machine learning TAA is!" when the released, assumedly not machine learning, FSR 3.1 is out at the same time.

Nothing "weird" about that "narrative" because it's what the users will be able to do even after FSR 3.1 release - since they won't be able to "upgrade" these old games to FSR 3.1 by themselves and will need to wait for developers to do that (which in many cases will never happen). DLSS and XeSS are drop-in DLL upgrades available for every game which support older versions.

 

[Subtlesnake]

Waiting like, maybe a week or two, for FSR 3.1 to have complete video of all the updates that were announced around the same time would've made more sense, what a weird decision.

AMD said FSR 3.1 would be out Q2, which ends in 2.5 months. Unless you have more up to date news?

 

[Cyan]

I think you will like today's DF video. 1.3 for sure advanced XMX and DP4a versions considerably.

my bad, I realised something today, some divine inspiration but I don't remember seeing it mentioned.

I downloaded the 1.3 SDK to try XeSS on Shadow of the Tomb Raider copying the libxess file, and I tested it using the .exe in the bin folder of the SDK.

And is it me or... doesn't XeSS support the equivalent of DLAA now? There is a native resolution setting where XeSS can be applied, unless I am wrong.

 

[Frenetic Pony]

AMD said FSR 3.1 would be out Q2, which ends in 2.5 months. Unless you have more up to date news?

Q2 means "anything between it was just posted right now and 2.5 months from now". If it's posted right now and I've not checked, that wasn't long to wait was it?

Also his entire thesis at the beginning of the video is "ML is so important to TAA!!!" A logical approach would be to prove this thesis, compare the new ML Plugins to the new non ML Plugin, then show evidence that the new non ML is still inferior by a significant margin. Unfortunately he took the "I will assume my hypothesis is true without testing, and now I will selectively choose evidence to prove it after the fact" approach instead.

 

[DavidGraham]

Also his entire thesis at the beginning of the video is "ML is so important to TAA!!!"

ML upscaling solutions have pretty much proven they are superior to FSR for several years already and they still continue to improve, but suddenly you expect FSR to match them based on what exactly?

AMD listed very specific improvements for FSR3.1 upscaling (for ultra performance mode), and in specific cases, we have yet to know if their improvements apply to other modes and other areas or not, so your best course of action is to wait and see, especially in light of their unknown release date.

 

[Subtlesnake]

Q2 means "anything between it was just posted right now and 2.5 months from now". If it's posted right now and I've not checked, that wasn't long to wait was it?

Also his entire thesis at the beginning of the video is "ML is so important to TAA!!!" A logical approach would be to prove this thesis, compare the new ML Plugins to the new non ML Plugin, then show evidence that the new non ML is still inferior by a significant margin. Unfortunately he took the "I will assume my hypothesis is true without testing, and now I will selectively choose evidence to prove it after the fact" approach instead.

But you were talking about holding the video a couple of weeks, which is completely arbritrary in the context of a 2.5 month release period.

Regarding the 2nd point, Alex was talking about the historical importance of ML when compared with other upscaling methods, in reference to the advances made by XeSS. I think that comment still stands even if FSR 3.1 completely catches up. And I notice the FSR 3.1 improvements listed by AMD make no reference to particle or transparency effects where ML solutions currently have a big lead.