AMD Announces Fidelity FX Super Resolution, their competitor to nvidia DLSS

[porina]

4 hours ago, xAcid9 said:

I reckon this will become like G-Sync vs Freesync.  AMD implementation is inferior but get widely adopted. 

We will have to wait for the quality comparisons, which is as much of a battleground as the performance numbers. Still, that's what I suspect will be the case. Apparently it will only use spatial scaling not temporal, so there is more risk of flickering for example.

 

2 hours ago, xAcid9 said:

Yeah, but i just found out AMD's published a patent about FSR recently and one image in particular show deep-learning involved so probably not a rebrand. 

I haven't seen the AMD presentation yet but Anandtech's writeup on it states they are not using machine learning based techniques. Also existence of a patent it not an indicator it will end up in a product any time ever. Companies file anything that is at all interesting as it may be of value even if they never use it themselves.

[LAwLz]

Sounds promising but I am very skeptical.

If I had to guess, it will not be anywhere near as good as DLSS currently is, but comparing it to DLSS is kind of moot. This is hardware agnostic and hopefully gets wide support (if it's good).

We can discuss DLSS vs Fidelity FX all we want, but that discussion will only matter when:

1) A game support both technologies and someone needs to make a choice which one to use.

2) When some fanboy wants to talk about how their favorite brand is superior to the competitor.

 

The reason why I am very skeptical is because the demo image they showed, running on a GTX 1060, looks quite frankly awful.

Left is native resolution, right is Fidelity FX in "quality mode":

 

If it's this bad then you might as well just run at a lower resolution and not use FidelityFX.

I seriously hope it improves before release, because if the comparison they posted is the current example they want to use to highlight how great it is, then it's in a VERY bad shape.

 

 

For comparison, this is an example of what DLSS looks like:

Left is native 4K, right is 1440P upscaled to 4K with DLSS:

[GreatnessRD]

I feel this is a crawl before you walk moment. AMD is in the crawling stage and over time, they'll be able to walk just as Nvidia did with DLSS 2.0. Consumers just have to do something they don't want to do, and that's give it... time. Hope it works well because that will be better for us all.

[Doobeedoo]

This will be fun to see, also being agnostic approach is good.

[Murasaki]

7 hours ago, HelpfulTechWizard said:

Time for rx400 owners to do a rx500 flash.

Guess as of tomorow ill have a rx580 4gb.

Got an RX480 so guess I gotta do that too... first timer on vbios flashing.

[xAcid9]

3 hours ago, porina said:

I haven't seen the AMD presentation yet but Anandtech's writeup on it states they are not using machine learning based techniques. Also existence of a patent it not an indicator it will end up in a product any time ever. Companies file anything that is at all interesting as it may be of value even if they never use it themselves.

Yeah, Tweaktown mentioned there will be no machine learning too, for FSR 2.0 maybe?

 

3 hours ago, porina said:

Apparently it will only use spatial scaling not temporal, so there is more risk of flickering for example.

But since it's not using temporal it will be less prone to ghosting unlike DLSS?

 

3 hours ago, LAwLz said:

The reason why I am very skeptical is because the demo image they showed, running on a GTX 1060, looks quite frankly awful.

Yeah, The foliage is blurred to the point it looks like 720p to my eyes. Might as well just drop the resolution to 1080p. 

 

 

[Stahlmann]

9 hours ago, BuckGup said:

I didn't watch the video but does it still rely on AMD training models for each game like Nvidia does?

DLSS 2.0 doesn't rely on training for specific games anymore. That's why it has been adopted by unreal engine as a developer tool. Or do you think Nvidia offers free Ai training for each indie dev that decides to put it in his game? The current DLSS is trained on a big library of images, not some specific game like DLSS 1.0 was.

 

AMD still has a long way to go to deliver an actual competitor to DLSS in terms of quality. It's impressive that they want to support older GPUs aswell and it looks like a great move in the current shortage that will likely last well into next year. But this is also likely a bottleneck. Remember, DLSS uses dedicated hardware "tensor cores". As it has been shown with many things in the past, dedicated hardware is pretty much always superior than relying on more "general" hardware like stream processors / cuda cores. But as AMD doesn't have these tensor cores or any equivalent hardware they have to work with what they got i guess.

 

Maybe this will also make Nvidia realease some form of "open" DLSS for non-RTX GPUs. Remember, G-Sync was also NVidia only and now is useable for AMD aswell because AMD pushed through with their open FreeSync.

 

I'm interested to see how this will work out.

[Bombastinator]

6 hours ago, RejZoR said:

Haha, running FSR on GTX 1060 was a power move indeed Also having FSR work on everything from Polaris cards and up is indeed a great way to extend their lifetime. I bet people who have RX 570/580 are already rejoicing and I even heard many RX 480 users deciding to flash them to RX 580 to get support. Also I can imagine there us quite some users who grabbed RX 5700 series back then as they were pretty capable. And now they'll be even more capable.

 

Given it's hardware agnostic, we can already expect wider adoption, because doing this means you've done work for AMD and NVIDIA. Doing DLSS means you've only done it for NVIDIA. Now it'll be just a matter of how much money NVIDIA is willing to throw at this situation to shill their tech. Seeing DLSS become hardware agnostic is something I'm not expecting at all.

My memory is far and away the most popular cards on steam are 1060s & 580s. If it runs on those this could be another freesynch situation.  Game developers will go for it because it can be sold better because the market is bigger.

[xAcid9]

2 minutes ago, Bombastinator said:

Game developers will go for it because it can be sold better because the market is bigger.

Not to mention it will be available to PS5, XBSX/S as well. 

 

6 minutes ago, Stahlmann said:

Remember, DLSS uses dedicated hardware "tensor cores"

Can 1660 series use DLSS as well? 

[igormp]

4 hours ago, porina said:

I haven't seen the AMD presentation yet but Anandtech's writeup on it states they are not using machine learning based techniques

They didn't, they mentioned that it doesn't rely on deep learning hardware:

Quote

but they are confirming that it doesn’t require any kind of tensor or other deep learning hardware. 

We still don't know for sure what tech is behind it, we will need to wait until the source is available.

[Stahlmann]

9 minutes ago, xAcid9 said:

Can 1660 series use DLSS as well? 

No. DLSS can only be used on RTX cards because it needs their "Tensor Cores" to function.

Only RTX 2000/3000 cards have these.

[xAcid9]

4 minutes ago, Stahlmann said:

No. DLSS can only be used on RTX cards because it needs their "Tensor Cores" to function.

My bad, I thought 1660 Ti have tensor cores. 

[Helpful Tech Witch]

1 hour ago, Murasaki said:

Got an RX480 so guess I gotta do that too... first timer on vbios flashing.

I can pm you if you need help.

Might take a couple tries though, depending on how much of the bios has to be rx580

[Murasaki]

3 minutes ago, HelpfulTechWizard said:

I can pm you if you need help.

Might take a couple tries though, depending on how much of the bios has to be rx580

Thanks! Won't do it right away though since I don't want to screw anything up and end up with no GPU... in these times it would be the worst thing.

[Helpful Tech Witch]

2 minutes ago, Murasaki said:

Thanks! Won't do it right away though since I don't want to screw anything up and end up with no GPU... in these times it would be the worst thing.

I completly understand. 

though its nice, on ati flash, yu can flash a bios to a card that has been bricked, if you have a igpu or second gpu.

[StDragon]

39 minutes ago, Stahlmann said:

DLSS 2.0 doesn't rely on training for specific games anymore. That's why it has been adopted by unreal engine as a developer tool. Or do you think Nvidia offers free Ai training for each indie dev that decides to put it in his game? The current DLSS is trained on a big library of images, not some specific game like DLSS 1.0 was.

AI training is just a fancy way of saying "optimized". It's just that it's not real-time and not performed on the client machine.

 

That last part is the problem IMHO. Both nVidia and AMD should provide APIs to game developers that let the client machine run the game in a demo loop to train optimization settings. This would reduce real-time load on the GPU while still allowing flexibility with future driver updates and not risk breaking compatibility and performance. It could also allow a wider adoption for game development. And who knows, maybe retraining again with a newer driver would yield more performance and improved image quality without having to change the game code.

[Murasaki]

2 minutes ago, HelpfulTechWizard said:

I completly understand. 

though its nice, on ati flash, yu can flash a bios to a card that has been bricked, if you have a igpu or second gpu.

Used my mobo's onboard graphics when my first 480 died so thats kinda assuring!

[decolon]

9 hours ago, xAcid9 said:

I reckon this will become like G-Sync vs Freesync.  AMD implementation is inferior but get widely adopted. 

likely will be the case bc dlss requires an rtx card iirc

[igormp]

1 minute ago, StDragon said:

AI training is just a fancy way of saying "optimized". It's just that it's not real-time and not performed on the client machine.

 

That last part is the problem IMHO. Both nVidia and AMD should provide APIs to game developers that let the client machine run the game in a demo loop to train optimization settings. This would reduce real-time load on the GPU while still allowing flexibility with future driver updates and not risk braking compatibility and performance. It could also allow a wider adoption for game development. And who knows, maybe retraining again with a newer driver would yield more performance and improved image quality without having to change the game code.

Running a demo loop for a couple hours on a specific game with subpar hardware wouldn't improve things, it might even make it worse.

You just don't update a model's weight in real-time, specially in such scenario where latency is critical.

[dizmo]

8 hours ago, GamerDude said:

DLSS is locked onto nVidia cards, and if you'd watched the vid, even GTX 1000 series cards can use FSR. Do you see nVidia doing anything resembling this? For sure, first gen FSR might be less than spectacular, so was DLSS 1.0 if you can recall. I'm just hoping FSR will be supported in the games I play, especially Metro Exodus PC EE (need a boost with RT enabled).

What's your point?

[Bombastinator]

41 minutes ago, xAcid9 said:

My bad, I thought 1660 Ti have tensor cores. 

My understanding is Any consumer Nvidia gpu that starts with a 1 has no tensor cores

[GamerDude]

30 minutes ago, dizmo said:

What's your point?

My point is, all this dissing on AMD's first attempt at DLSS-like super res, it'd take time for AMD to get better at this, much like how DLSS 1.0 was less than spectacular at its launch, DLSS 2.0 is awesome now. nVidia had a head start, plus locking DLSS to its own hardware ensures that it'd work better, (hardware + software > software only) I'm not arguing that. 

 

This works with both older AMD GPU's and GTX 1000 cards, no reasonable person would expect FSR to match DLSS, especially in its first iteration when AMD is working out the kinks. Here, and in other forums, nVidia boys harp on the PQ, not even bothering to take note that it works on more recent nVidia GPUs as well. 

 

Whatever the case, I'll ignore this thread as I don't wanna get involved in any arguments.....

[RejZoR]

I also wonder what will be resolution limitations. A lot of games refuse to enable DLSS if you're using "only" 1080p. For example Shadow of Tomb Raider refused to give any control of it at 1080p. When i had 1080p monitor I had to use DSR to run game at 1440p and then "downscale" it with DLSS. Or now that I have 1440p monitor it gets available by default. I wonder what limitations FSR has in this regard...

[Stahlmann]

1 hour ago, StDragon said:

AI training is just a fancy way of saying "optimized". It's just that it's not real-time and not performed on the client machine.

 

That last part is the problem IMHO. Both nVidia and AMD should provide APIs to game developers that let the client machine run the game in a demo loop to train optimization settings. This would reduce real-time load on the GPU while still allowing flexibility with future driver updates and not risk breaking compatibility and performance. It could also allow a wider adoption for game development. And who knows, maybe retraining again with a newer driver would yield more performance and improved image quality without having to change the game code.

These algorythms to optimize DLSS need to run in a big datacenter. Running it on some client machines for a couple of hours or weeks won't help.

 

Seeing how well it works, with sometimes even improving image quality compared to native resolution, i think the current approach is a good one.

[Stahlmann]

9 minutes ago, RejZoR said:

I also wonder what will be resolution limitations. A lot of games refuse to enable DLSS if you're using "only" 1080p. For example Shadow of Tomb Raider refused to give any control of it at 1080p. When i had 1080p monitor I had to use DSR to run game at 1440p and then "downscale" it with DLSS. Or now that I have 1440p monitor it gets available by default. I wonder what limitations FSR has in this regard...

That's because it was DLSS 1.0.

 

The 2.0 version does not need any specific target resolution.

It works with all standard resolutions, including ultrawide.