AMD FSR antialiasing discussion

[DavidGraham]

ps. DLSS launched with just one game too. XeSS will most likely launch with just one game. Pretty much every new tech out there starts with just one game.

DLSS 1 required massive efforts and work, DLSS 2 doesn't, so it launched with dozen of games, FSR 2 is launching after FSR 1, and only managed ONE measle DLSS 2 enabled game. The game was previewed months in advance, yet officially launched later.

FSR 2 hand crafted approach is even more time consuming than DLSS 2, it needs it to look comparable to DLSS 2, otherwise, it's not gonna look good at all.

They actually never said that.

Nope they flat out did.

 

[madhatter]

That's up for debate, they never said that, some interpreted what they said like that. Only thing for sure is performance difference and on their graphs, while DP4a seems to take about twice the time of XMX version, it's still just a fraction of overall frametime (and the graph doesn't actually have any scales included so that "twice the time" is just a guess based on bar length)
edit: The exact quote is "with smart quality performance trade off" which could just as well mean you need to compromise some quality to reach same performance, not that you couldn't get same quality period.

Mhm, thanks for clarifying. It won't be too long until we get to see it in action, hopefully.

At least I'm still steering clear of any scalers so long as I can reach my desired performance by lowering settings instead of resolution

You don't think they can be preferable in certain games?

None of these bosos are worth the time, they said FSR 1 Ultra Quality was comparable to DLSS Quality many times in the past, now they claim FSR 2 is a HUGE improvement, and still comparable to DLSS 2, oh really?

Computerbase seems to have reached similar conclusions, if you trust them more.

 

[Picao84]

It's my understanding, and someone can correct me if I'm wrong, that training requires higher precision computing that is performed in massive offline clusters generally. Tensors are good for inferencing only, that's their primary purpose. Very fast lower precision matrices searches.

Ampere has High Precision Tensor Cores for training.

https://developer.nvidia.com/blog/accelerating-ai-training-with-tf32-tensor-cores/

 

[techuse]

DLSS 1 required massive efforts and work, DLSS 2 doesn't and launched with dozen of games, FSR 2 is launching after FSR 1, and only managed ONE DLSS 2 enabled game.

Nope they flat out did.

Link?

 

[OlegSH]

It's my understanding, and someone can correct me if I'm wrong, that training requires higher precision computing that is performed in massive offline clusters generally.

Training can live up with as low precision as FP8 for certain layers in transformers, BF16 is quite common, TF32 can be used virtually anywhere -- usually higher precision is used for fully connected layers with sigmoid activation.
Inference can go below 8 bits thanks to quantization, but at the expense of prediction precision most often (lower quality).

 

[Kaotik]

DLSS 2 doesn't and launched with dozen of games

Just double checked, they had 4 games in the launch article (Control, MW5, Wolfenstein, Deliver us the Moon), but I'm pretty sure one of them (Control?) got support clearly before the others.

FSR 2 is launching after FSR 1, and only managed ONE DLSS 2 enabled game.

FSR 2.0 has nothing but the name to do with FSR 1.0. DLSS 1.x and 2.x are closer to each other engine implementation wise.

You don't think they can be preferable in certain games?

So far the artifacts are annoying me more than dropping something from ultra to high or even medium. If they come up with artifact free solutions, I might think again.

 

[snc]

Very impressive, surprise so close to dlss2, waiting for other games

 

[DavidGraham]

Just double checked, they had 4 games in the launch article (Control, MW5, Wolfenstein, Deliver us the Moon), but I'm pretty sure one of them (Control?) got support clearly before the others.

Many more followed shorlty after that, with very rapid successions and iterations.

FSR 2.0 has nothing but the name to do with FSR 1.0. DLSS 1.x and 2.x are closer to each other engine implementation wise.

AMD claimed days to implement FSR 2 in DLSS games, yet they only launched ONE, took months to do it, and still relied on the basis of DLSS to even deliver it, clearly time is not their friend at all.

 

[Malo]

Training can live up with as low precision as FP8 for certain layers in transformers, BF16 is quite common, TF32 can be used virtually anywhere -- usually higher precision is used for fully connected layers with sigmoid activation.
Inference can go below 8 bits thanks to quantization, but at the expense of prediction precision most often (lower quality).

Thanks, good to know. I guess the main point is that training has nothing to do with real time gaming. You don't train during gameplay.

 

[trinibwoy]

So now do you who disliked dlss, want to use fsr now?

Probably since some folks dislike Nvidia more than they dislike DLSS. Fact is image reconstruction is everywhere now. We need some form of cheap temporal integration to squash annoying shader aliasing and that’s where FSR, DLSS, TAAU, XeSS come in.

 

[OlegSH]

You don't train during gameplay.

Of course, unless that's stuff like this, but this stuff is completely unrelated to the topic.

 

[DegustatoR]

I'm not sure why people are surprised by initial FSR2 showing. Epic's TSR is basically the same thing and it was already compared in Ghostwire with DLSS with rather similar results - meaning that a modern TAAU it's pretty close to DLSS on static images and is a bit worse in motion. Same was bound to be true for FSR2 so why the weird reaction?

DLSS will still provide better quality and will still be able to do so with lower base resolutions - Performance preset is 100% usable with DLSS for example.

Otherwise FSR2 is what FSR1 should have been - a great state of the art TAAU solution available to everyone now, actually capable of competing against other options.

 

[Kaotik]

Many more followed shorlty after that, with very rapid successions and iterations.

And AMD has 11 games listed already to follow with FSR 2.0

AMD claimed days to implement FSR 2 in DLSS games, yet they only launched ONE, took months to do it, and still relied on the basis of DLSS to even deliver it, clearly time is not their friend at all.

3 days was the optimal situation, there may be more into it in the beginning (when were the tools really available etc?). Of course first game is one where there's less work to be done, it would be stupid not to take advantage of it.

 

[DavidGraham]

3 days was the optimal situation, there may be more into it in the beginning (when were the tools really available etc?). Of course first game is one where there's less work to be done, it would be stupid not to take advantage of it.

Again, only launched one that took several months to implement, if it was easy as they claimed, they would have launched with more games, and took far less time to implement FSR 2 in this one game.

And AMD has 11 games listed already to follow with FSR 2.0

In 3 days? 3 weeks?

 

[Cat Merc]

DLSS 1 required massive efforts and work, DLSS 2 doesn't, so it launched with dozen of games, FSR 2 is launching after FSR 1, and only managed ONE measle DLSS 2 enabled game. The game was previewed months in advance, yet officially launched later.

FSR 2 hand crafted approach is even more time consuming than DLSS 2, it needs it to look comparable to DLSS 2, otherwise, it's not gonna look good at all.

Nope they flat out did.

The certainty you speak with is quite funny.

Deathloop is the first because they were AMD's partner for creating the dev pipeline. Arkane was their main partner for building the API others will use. This was in the GDC presentation.

 

[Granath]

Again, only launched one that took several months to implement, if it was easy as they claimed, they would have launched with more games, and took far less time to implement FSR 2 in this one game.
In 3 days? 3 weeks?

Why do you hate everything that is not Nvidia branded ?

 

[PSman1700]

Can we have less of this please? For the record my rig rocks an RTX3070 and I've used DLSS with no problems. Just that, a software developer, I refuse narratives where a software solution mandates a very specific proprietary piece of hardware for something that is highly generic in nature.

Well, i dont know if it was you either, didnt name anyone, i ask those who where saying they didnt want to have anything to do with DLSS because of smeary and blur, and hell if it was used in benchmarks. That narrative seems largely changed now when TAAU-like solutions have arrived to the PC aswell.

Probably since some folks dislike Nvidia more than they dislike DLSS. Fact is image reconstruction is everywhere now. We need some form of cheap temporal integration to squash annoying shader aliasing and that’s where FSR, DLSS, TAAU, XeSS come in.

I think it has not much to do with them being against NV at all. They are fighting the wrong fight anyway, FSR2.0 isn't for them to begin with.

I'm not sure why people are surprised by initial FSR2 showing. Epic's TSR is basically the same thing and it was already compared in Ghostwire with DLSS with rather similar results - meaning that a modern TAAU it's pretty close to DLSS on static images and is a bit worse in motion. Same was bound to be true for FSR2 so why the weird reaction?

DLSS will still provide better quality and will still be able to do so with lower base resolutions - Performance preset is 100% usable with DLSS for example.

Otherwise FSR2 is what FSR1 should have been - a great state of the art TAAU solution available to everyone now, actually capable of competing against other options.

I have no idea at all, TAAU solutions like FSR2 have been available for awhile now, not just in UE5 but also other engines. On consoles it has been even more present, better at it. The whole idea of hand-coded temporal reconstruction is a different beast to the automatic/ML one. TSR, TAAU, insomniacs variant etc have always been very close to what DLSS was doing.

 

[Cat Merc]

I'm not sure why people are surprised by initial FSR2 showing. Epic's TSR is basically the same thing and it was already compared in Ghostwire with DLSS with rather similar results - meaning that a modern TAAU it's pretty close to DLSS on static images and is a bit worse in motion. Same was bound to be true for FSR2 so why the weird reaction?

DLSS will still provide better quality and will still be able to do so with lower base resolutions - Performance preset is 100% usable with DLSS for example.

Otherwise FSR2 is what FSR1 should have been - a great state of the art TAAU solution available to everyone now, actually capable of competing against other options.

TSR has quite terrible ghosting and the end presentation always seemed way too blurry in motion. I never understood why people seemed to like it so much. Ratchet and Clank's TAAU is perhaps the only one I would say is great?

It really wasn't a given that FSR 2 would end up as well as it did.

 

[DavidGraham]

Deathloop is the first because they were AMD's partner for creating the dev pipeline. Arkane was their main partner for building the API others will use. This was in the GDC presentation.

No, none of that is in the GDC presentation.

Deathloop is a DLSS enabled game that took months to properly implement FSR 2, and didn't allow for any other DLSS enabled games to be released with FSR 2 alongside it, despite AMD claiming otherwise.

 

[Cat Merc]

No, none of that is in the GDC presentation.

Deathloop is a DLSS enabled game that took months to properly implement FSR 2, and didn't allow for any other DLSS enabled games to be released with FSR 2 alongside it, despite AMD claiming otherwise.

That was one of two presentations.

But please do go on, I do get a chuckle from this.