AMD FSR antialiasing discussion

[Krteq]

Is it just me or Alex completely ignored FPS increase benefit in that review?

And in that KingsHunt comparison with TAAU - he compared it with performance preset only (which makes sense considering both uses same input resolution) while FPS measurement or comparison with other preset is completely missing

Good question. You can do this with a combination of DLSS and DSR today. On a 1080p monitor upscale to 4K using DLSS and downscale back to 1080p using DSR.

Yes, it's possible to use VSR in combination with FSR.

 

[Rootax]

Is it just me or Alex completely ignored FPS increase benefit in that review?

And in that KingsHunt comparison with TAAU - he compared it with performance preset only (which makes sense considering both uses same input resolution) while FPS measurement or comparison with other preset is completely missing

Yes, it's possible to use VSR in combination with FSR.

There is a segment about performances and gpu utilisation. But yeah the main focus was on the tech and the PQ it seems.

 

[Leoneazzurro5]

It's quite sad when it's mentioned that some devs using UE4 don't even know about TAAU.

Do you really think that's the case? What's the chance that MOST people actually MAKING games and working on a daily basis with the people who created the engine woud not know about it? If it not widely used, quite probably there are downsides to it.

 

[Flappy Pannus]

Is it just me or Alex completely ignored FPS increase benefit in that review?

He shows the performance delta between it in performance mode and UE4's TAAU, both of which are scaling from 1080p. The difference in GPU % was basically identical - considering that UE4 TAAU is reconstruction and not just 'upscaling' and delivered a far better (imo) result, not exactly a point in FSR's favour.

 

[Cyan]

great review, as usual. In a similar fashion -and conclusions-, Tom's Hardware also mentioned the big advantages and the shortcomings, adding that Terminator Resistance looks better with FSR on at ultra quality than native 4k.

https://www.tomshardware.com/uk/news/amd-fidelityfx-super-resolution-fsr-tested

They even tried it on a nVidia MX330 and it worked!

That being said, those reviews miss an important point for different people than us who just want to play. They dont care about the image quality being perfect nor counting pixels, but if their games run more smooth, their GPUs consume less energy..., that's all that matters and where nVidia faces a tough competition.

The tech is going to be on consoles for mass adoption. Imho, nVidia could also lose this one -as they lost the GSync vs Freesync battle-.

 

[Cyan]

also, AMD shared this screengrab showing the performance of a GTX 1060 with this technology.

The GT 1030 can be a decent competitor at 1080p with this technology.

 

[Cyan]

another review on FSR with interesting conclusions

 

[Davros]

Fixed the AMD slide

 

[digitalwanderer]

Are there any games actually out that don't suck that I can test this on? I'm not a big fan of any of the titles they presented, well I don't even really know them.

What is the best game to test this out on? I gotta play with it!

Also, 22nd Racing whatever isn't out yet. I am a bit annoyed, but then again this morning I "helped" the vets trim my dogs nails and just got back.

I'm the one in the middle under Buffy, forgive the off-topic but it's almost cute and it WAS stressful!

 

[Mindtaker]

So I tested it a bit. The main differences on terms of image quality comes vs Native + CAS comparison.

1440p FSR Ultra quality

1440p native without CAS

1440p native with CAS

 

[PSman1700]

looking at Alex's more capable analysis that the rest of these amateurish websites there is a clear and significant downgrade in details. Its not little and hard to notice. Its a visibly softer image with compromised details in multiple areas. Alex also points how the aliasing of the base resolution remains in the upscaled image. So pixel crawl, shimmering and all that stuff remains. Image stability was one of the highlights of dlss in a game like Death Stranding for example

Digital Foundry's video here:

Watching that first, then coming into this thread was somewhat surprising when seeing some of comments from people that were generally impressed. If anything I've been generally upbeat on this before release, I've argued that I don't think this needs to 'match' DLSS to be worthwhile at all as some have, simply by nature of it being open source and supported across a wide variety of hardware is a feature unto itself - I definitely don't want it to replace DLSS as the gold-standard, but having another 'decent' reconstruction tech that's easily integrated as another option for non-RTX cards is a good thing, especially as on the PC we see a far sparser use of checkerboarding/TAA upsampling across games.

However, the results as Alex showed are less than impressive, and bear in mind Alex doesn't even make a direct comparison to DLSS in the video (DLSS isn't even mentioned until the very end). Rather just comparing it against regular bilinear and TAA upscaling, it falls down hard against the latter imo.

UE4 TAA looks significantly superior here. I can't speak to how good it is compared to say, Insomniac's very impressive reconstruction tech or basic checkerboarding, but I don't expect it would far well against those either. So that this is at 'least' a worthwhile method for the new consoles to squeeze out extra performance is dubious when it likely can't compete with already established reconstruction methods they've been using for years.

So what is this for? Based on what I've seen so far, if I see this - and only this - in a game as the only reconstruction method option, that will basically be its purpose - 'better than nothing' - and perhaps compared to bilinear upscaling + sharpening, even that meagre improvement might be largely negated. Of course goes without saying if this can't compete with regular TAA upscaling it's not remotely in the same league as DLSS.

Maybe it will improve, but as others have postulated I don't see that happening until it becomes more than a spatial upscaling method. From what I've seen it's less FSR and more BBB (Barely Better Bilinear).

I guess after this less than impressive showing the question is: was this already being researched or in the pipeline for AMD or was it a knee jerk reaction quickly put together just to have something, anything, to appease fans asking for a DLSS competitor? Sounds like a terrible waste of time..

Thanks for the technical insight. Having seen all the videos, comments, and reviews, FSR is actually less impressive then it was thought to be.

If Alex's video focused on DLSS as a comparison

That wouldnt be all that fair, their not the same technologies. DLSS is a reconstruction tech, FSR is just an upscaler, like DF mentioned. FSR isnt competing with DLSS, its competing with other upscaling techs.
Kinda sad for the consoles missing the boat on both ray tracing and reconstruction tech by going with RDNA1/2. My guess is RDNA3+ will fare much better.

 

[Rootax]

Are there any games actually out that don't suck that I can test this on? I'm not a big fan of any of the titles they presented, well I don't even really know them.

What is the best game to test this out on? I gotta play with it!

Also, 22nd Racing whatever isn't out yet. I am a bit annoyed, but then again this morning I "helped" the vets trim my dogs nails and just got back.

I'm the one in the middle under Buffy, forgive the off-topic but it's almost cute and it WAS stressful!

View attachment 5592

Nice guard here !

 

[Rootax]

Thanks for the technical insight. Having seen all the videos, comments, and reviews, FSR is actually less impressive then it was thought to be.



That wouldnt be all that fair, their not the same technologies. DLSS is a reconstruction tech, FSR is just an upscaler, like DF mentioned. FSR isnt competing with DLSS, its competing with other upscaling techs.
Kinda sad for the consoles missing the boat on both ray tracing and reconstruction tech by going with RDNA1/2. My guess is RDNA3+ will fare much better.

Consoles are doing ok with RT right now...

 

[DegustatoR]

The main differences on terms of image quality comes vs Native + CAS comparison.

You should try adding CAS via injection instead of relying on game's own implementation - which seems to do quite a bit of oversharpening.

 

[PSman1700]

Consoles are doing ok with RT right now...

Exactly, their doing what was being done in 2018, if even that. Thats considered ok, even by some pc gamers, but it would have been something if they released them just one year later, it wouldnt have done all that much of a difference seeing the supply problems, scalpers etc.

 

[see colon]

great review, as usual. In a similar fashion -and conclusions-, Tom's Hardware also mentioned the big advantages and the shortcomings, adding that Terminator Resistance looks better with FSR on at ultra quality than native 4k.

https://www.tomshardware.com/uk/news/amd-fidelityfx-super-resolution-fsr-tested

They even tried it on a nVidia MX330 and it worked!

That being said, those reviews miss an important point for different people than us who just want to play. They dont care about the image quality being perfect nor counting pixels, but if their games run more smooth, their GPUs consume less energy..., that's all that matters and where nVidia faces a tough competition.

The tech is going to be on consoles for mass adoption. Imho, nVidia could also lose this one -as they lost the GSync vs Freesync battle-.

This is going to extend the life of midrange gaming laptops. Pretty happy about that, honestly.

 

[snarfbot]

Seems pretty awesome from what I've seen. Native falling somewhere between quality and ultra quality. Essentially get a free bump up in res.

 

[Kaotik]

I wonder why it needs driver support.

Could be as simple as "game support" in drivers in general

Digital Foundry's video here:

Watching that first, then coming into this thread was somewhat surprising when seeing some of comments from people that were generally impressed. If anything I've been generally upbeat on this before release, I've argued that I don't think this needs to 'match' DLSS to be worthwhile at all as some have, simply by nature of it being open source and supported across a wide variety of hardware is a feature unto itself - I definitely don't want it to replace DLSS as the gold-standard, but having another 'decent' reconstruction tech that's easily integrated as another option for non-RTX cards is a good thing, especially as on the PC we see a far sparser use of checkerboarding/TAA upsampling across games.

However, the results as Alex showed are less than impressive, and bear in mind Alex doesn't even make a direct comparison to DLSS in the video (DLSS isn't even mentioned until the very end). Rather just comparing it against regular bilinear and TAA upscaling, it falls down hard against the latter imo.

UE4 TAA looks significantly superior here. I can't speak to how good it is compared to say, Insomniac's very impressive reconstruction tech or basic checkerboarding, but I don't expect it would far well against those either. So that this is at 'least' a worthwhile method for the new consoles to squeeze out extra performance is dubious when it likely can't compete with already established reconstruction methods they've been using for years.

So what is this for? Based on what I've seen so far, if I see this - and only this - in a game as the only reconstruction method option, that will basically be its purpose - 'better than nothing' - and perhaps compared to bilinear upscaling + sharpening, even that meagre improvement might be largely negated. Of course goes without saying if this can't compete with regular TAA upscaling it's not remotely in the same league as DLSS.

Maybe it will improve, but as others have postulated I don't see that happening until it becomes more than a spatial upscaling method. From what I've seen it's less FSR and more BBB (Barely Better Bilinear).

You ignore all but the Performance mode - the one mode AMD actually says you shouldn't even use.

 

[snc]

very underwhelming results (worse results than unreal eninge taa upsample), maybe quality will improve in future

 

[Cyan]

One thing i wish alex mentioned was how games look when compared to the native resolution they are being upscaled from rather than native 4k. Is the improvement noticeable with fsr upscaling from 1612p versus a native 1612p? What are the performance trade offs of that ect

I didnt go in expecting fsr to be as good as native 4k but i was expecting comparisons from basic 1440p to fsr upscaled from 1440p ect.

great point.

This is going to extend the life of midrange gaming laptops. Pretty happy about that, honestly.

hmmmm never thought of that tbh, but now that you mention it, imagine the benefits. Laptops not heating up to 90ºC anymore, while having a somewhat decent image quality to play with and the possibility to enjoy games which were almost unplayable.

As for desktop computers and from personal experience, the thing I care the most about a future GPU -which I am going to need in a year or two- is power efficiency to play at 1440p, adding robust framerate. So I am torn between nVidia, AMD or Intel and have time to decide.

The thing I care the most is the electricity bill and the temperature of the components. That's why I always preferred the most efficient hardware rather than brute force hardware. In that sense, oddly enough, AMD has the best GPUs, imho, as of now