I'm still so amazed that we can manage to produce a nicer looking image at a LOWER resolution while GAINING performance. Absolutely blows my mind, honestly. I love tech videos like this so much!!
Yep I love that. Such a massive win. I'm assuming this will be used in console games then too, so something like Horizon they could run at 1440 FSR'd up to 4k and giving the devs some extra power left over? (Or at the very least guaranteeing a locked 60)
@@clownavenger0 That's becoming true for more games, to the point where it makes sense to use DLSS even just for a higher image quality in addition to performance boost.
Bottom line: If you have an RTX GPU, use DLSS. If you have anything else, you have FSR 2.0. Definitely a win for AMD and for every gamer without an RTX card.
@@tazogochitashvili6514 lmao yes Nvidia doesn't give a fuck about its customers amd doesn't care too much either since both only look for money Ofc but AMD takes the role of the good cop and nvidia is the ahole cop f'ing over everyone that hasn't bought a product of theirs in the last 2 years Edit: and also those that buy their products regularly, so basically everyone
FSR benefits everyone. There will likely be developers that choose to invest into implementing FSR, but not DLSS since it will benefit a smaller proportion of their player base. This is an absolute win for gamers, thank you AMD (BTW I only have nvidia cards)
Big improvement? We rarely play games in still image. This has artefacts in motion which is a big no no. This will make the overall image look way shimmery in motion much like HFW performance mode.
@@Downloadfreak It's not. Did you see the breaking apart of image in motion at performance mode reconstruction? Shimmering extravaganza! How is that an improvement? Audio and visual comprehension much? Lol.
@@jacobartzavia6096 Exactly! Also the lower memory bandwidth on Series S will do better with an FSR-upscaled image from 1080p/720p rather than a native 1440p in games.
I love that DLSS gets rid of that awful shimmering in so many games and it's nice to see FSR should at least improve things in many cases too. For some reason when I see objects like fences doing that in games it has always bugged me.
@@hi_tech_reptilez I would say transparency is a much bigger challenge. Even with raytracing and 3090 games need tons of hacks to get away with faking it.
@@hi_tech_reptilez Personally I don't mind Aliasing as much as motion blur, with the lack of progress on non temporal AA solutions ghosting and blur has become the biggest problem with modern gaming for me as often there is no way to turn TAA off in the settings menu.
They probably did that intentionally so that in side by side comparisons with DLSS the details on geometry etc look sharper. Hopefully they tune the default down once they're not just trying to prove themselves.
Yeah I was thinking that, 10 seems a bit too high which could be causing a few of the other problems, it's a shame Ales didn't show some of those areas with the sharpness set to 7, 5, and 3 to see the results.
As someone who works with FSR 1.0's code a lot, and this has the same sharpness pass I have always found that 6-9 was the sweet spot, never used a game where I had to go lower or higher (unless the games TAA was a super blurry mess however this replaces the games AA) so 10 should realistically never be required. If you can't be bothered to test for the best value just set it to 8 and forget it though, that's my tip.
Seeing as DLSS will never come to current gen consoles, this is a huge step up for them compared to native rendering/gen 1 upscaling. Of course, there’s also UE5’s TSR on consoles, but that will be restricted to UE5 games. Would love to see a comparison video of MS flight sim after FSR 2.0 is implemented.
@@puffy3146 DLSS is exclusive to Nvidia cards, specifically gpus’s with tensor cores/machine learning (RTX 2060 and above). Current gen consoles use AMD hardware that is not compatible with DLSS.
@@puffy3146 I think we’ll start seeing it used more on console beginning in 2023 but it’s up to individual studios to implement it so have to wait and see.
Amazing technology. How did someone come up with this stuff, it’s so bloody complicated. Bravo to AMD for making it applicable across competitor GPU’s.
This is the natural evolution of TAA, where you attempt to use the additional information from previous frames to improve the output detail. This is "just" retaining more useful data so you can usually increase the output resolution as well as detail. The "just" is that this is actually horribly complicated to do without artifacts like "slime trails", strobing, or shimmer - like most things, the idea is a lot easier than the work!
The most competent analysis I've read since the FSR 2.0 release (only yesterday)! Such a good job, in such a short amount of time. Anyway the gap between FSR 1.0 and 2.0 is huge and it will grow for sure with the next updates!
This is pretty interesting. AMD have made some nice advancements with FSR, hopefully the ghosting can be patched out. Either way, it's nice to see more competition in this area of rendering.
by it's very nature temporal based AA will always by its nature create ghosting. It's why MSAA always looks nicer as its mid process and actually takes 3D information into account instead of just the depth buffer.
The actual pixelation you can sometimes see seems to also be partially caused by AMD's implementation to avoid ghosting (i.e. discarding the history buffer for occluded pixels, something you can see happening with the sidemovement with the antenna).
I have a feeling FSR 2.0 will be like DLSS is, where the quality of it's use is on a per-game basis, some do it better than others, so yeah, like Alex said, we need to see more games with FSR 2.0 but so far, the results look really promising and it's likely only going to get better with future revisions of FSR.
Yes, I'm familiar with different methods of AA and how they're implemented. However, as was noted in the video, this is just ONE implementation of FSR2.0, in a single game. Time and further development will tell how this plays out in in other games.
@@Draggobuttboi , There ARE solutions to fix the ghosting that TAA tends to cause. You can analyze the surrounding pixels and provide an upper limit on brightness for the aliased part. Similarly there are ways to use the motion vector values to apply motion compensation. So it's NOT an unfixable problem. I'm not sure how much VRR and/or Dynamic Resolution affects these workarounds.
Extremely interesting. Excited to see this on Microsoft Flight Simulator and Forspoken. Was extremely lucky to find a 6800xt at MSRP in 2020, but I have been missing DLSS. Happy there's competition now and that it's open sourced.
Super informative dive into the various stress tests for the reconstruction techniques. FSR 2.0 looks exciting. Not as robust as DLSS, but the broader hardware applicability and performance savings in quality mode are great to see. I'm hoping for some future integration into XSX and PS5 to hit 4k60 from either 1080p or 1440p internal resolutions in games with engines that would otherwise not have such a good reconstruction techniques available to them
Wow, this is an absolute game changer. Especially in the console space. With this- the PS5 & Xbox Series X can now conceivably run every title at 4K/60fps and with an image quality that is arguably better than native. On the PC side it’s great that this also works on any GPU. Cant wait to test it out myself
@@LouSassol69er Sadly most don’t. I’ve only had a DLSS capable GPU for around 6 months and the tech is incredible. It’s astounding to me that this isn’t implemented in every single game. With FSR 2.0 there is no longer an excuse- because you don’t even need an RTX card or dedicated tensor cores, it just works on everything.
A lot of console games were probably already doing something like this internally. But many temporal upscaling solutions are clearly more rudimentary than this, it's good to have an open source solution that everyone can use, and presumably submit patches back to.
The lateral test zoomed in on the device was the most interesting. TAA seems to be ghosting a lot as it is weighting samples too evenly such that stale data is still present and contributing strongly to pixel color, DLSS seems to be more cleverly weighting its samples by giving stale samples much lower influence on pixel color but if a sample is that stale it should probably just be evicted, its still ghosting just not as badly as TAA, meanwhile FSR is taking an aggressive stance on ghosting by evicting stale samples with prejudice this removes the trail of the device, BUT that means the recently disoccluded grass is now noticeably under sampled making it noisy and pixelated.
As a gtx 1070 scrub, this is great news. Can’t wait for fsr 2.0 to hit dying light 2 as I found the fsr 1.0 implementation in that game to be a very poor solution
Finally. It bugged the crap out of me that people kept comparing FSR to DLSS when they weren't trying to do the same thing at all. Now that comparison actually makes sense.
I'm just waiting for all this tech to be used as normal anti aliasing. I'm sick of the super soft look of TAA and ghosting in motion. Because these upscalers are designed to be as crisp and 'sharp' as possible, I'm guessing they'd produce significantly better results than TAA if they could use native resolution.
@@aweigh1010 it can certainly be implemented pretty well, a lot of frostbite games look quite good with it, but you only need to play halo infinite or almost any recent ue4 game to realise it can have serious issues. Anyway, aside from whether you prefer it over older methods, it would be ignorant to completely ignore its flaws. Flaws that dlss and fsr 2.0 do a decent job of avoiding, even with their lower resolutions.
@@FarmedRice haven't personally played doom, but I don't doubt that great TAA Implementations can exist, but if there's going to be a push to include dlss or fsr in newer games, it would be nice if it came along with a native resolution option too. Just for all those games out there with worse TAA than doom
1.05 ms added for a chip without dedicated silicon is simply awesome. I wish that we had 2022 image processing technologies in 2003. Would have made the transition to flat panels so much less annoying back then.
I’m not very tech savvy so I didn’t understand the added ms. What does that mean in terms of FPS? So if a game’s running at 60fps and we turn on FSR 4K Quality mode, what can we expect the FPS to be abouts?
@@TheHulksMistress Every frame takes time to render, this is often measured in milliseconds. I think 16.6 milliseconds per frame is needed to hit 60 fps. The added milliseconds is how much extra time is taken to render the effect. For example FSR 2.0 may take an extra 2ms to render, however, lowering the resolution by so much saves more time than it adds. Therefore you get more frames, so if a game is running at 60 at native 4k, and you turn on FSR 2.0 quality, you almost certainly will be above 80fps or higher.
That might actually get better with driver and FSR optimizations but for the start, it's already in a good state. It will be interesting to see with driver and FSR optimizations if that will translate to all FSR 2.0 games or if it's on a game per game basis.
@@Thyispro great, thanks for the example. I love that. I’ve watched enough Digital Foundry videos to know what devs can do with an extra 20fps worth of power to play with
Speaking of LCD's..I remember my dual 19" Princeton Graphics CRTs' back in the day XD...The model was Princeton Ultra 97. 1600 x 1200 resolution on both monitors...of course at the time I had a GT 9500 back then.
I wouldn't be surprised if FSR 2.0 started showing up as games as the default setting, in PC and consoles with it looking as good as it does and improving performance so much at the same time.
Maybe longer term but I doubt that will happen for a few years, but from a developer's point of view, having that on as default would lower the hardware requirements needed, that could mean more game sales.
Sony has to add it to their sdk so God only knows how long it's gonna take those useless fucks. We got vrr 3 yrs later have yet to get texture filtering for every game like Xbox does and 120hz. So yeah good luck with that.
Always good to see your extremely thorough testing, finding issues that no other outlets seem to notice. It's certainly a huge improvement over FSR 1.0 that will be very welcome for many, but I'll definitely continue using DLSS when I can. One thing you didn't touch on that's been talked about quite a bit though - anti-aliasing quality on slanted lines. Many people seem to be under the impression that FSR 2.0 actually handles this better, though it seems like about a tie to me. I'd be interested to hear your take on that, Alex.
I was watching a second time and I noticed that he also does not take segments where DLSS suffers from ghosting, because generally DLSS suffers from ghosting, and that happens in this game as well. One of the big surprises in all the video comparisons so far was that FSR 2.0 handles ghosting way better than DLSS, yet here in this video Alex almost succeeded in making it look the other way around 14:41
@@ctrlectrld Well different test cases will inevitably provide different results. It is odd that he didn't seem to produce the DLSS weapon sight ghosting that other reviews have noted (but also didn't say anything about the severe grainy trail on the grass on FSR 2.0 in the ghosting test.) I do wonder if some of these issues might be caused by the sharpening pass though and will look better with it turned down a bit.
@@nimbulan2020 Yep, really weird. He wrote on twitter that sharpening is the reason for the grainy trail. Thing is, people noticed the grainy trail, because it's very noticeable, and that's maybe why he didn't say anything about it, cause he didn't need to. But again, this way I am not sure he's presenting things fairly. As things stand, if we only had this video as reference, we would come to the conclusion that FSR 2.0 handles ghosting worse than DLSS, which is quite unfortunate.
@@ctrlectrld Well is it any less fair than other reviews not even mentioning the very noticeable issues that FSR 2.0 has? Realistically, reviewers have a limited about of time to test things, and they're not always going to catch every issue or be able to test every configuration. Alex stuck with the default sharpening setting since that's the one people will most likely use. It may look better in other games, we'll have to see when the time comes.
@@nimbulan2020 It seems to me that other relevant reviews, like the one by Hardware Unboxed, show the good and the bad. But in my previous comment I was referring to the ghosting issue specifically; in this case I feel that the topic is not being addressed fairly since ghosting has always been a main issue in the temporal domain, so much that DLSS is still struggling with it to this day. AMD with FSR 2.0 wanted to tackle specifically this issue and came up with a clever solution utilizing the depth buffer to reduce ghosting artifacts. It does a good job, at least better than DLSS, and it's one of the main perks of FSR 2.0. You can't just leave this whole thing out from a specialized in depth tech review, and on top of that even pick a segment where the better solution looks worse. It's not a fair and complete representation in my opinion.
@@DanKaschel $700 1080ti 2017 is one of the best purchase I made. This card is still decent today, especially if fsr 2.0 is coming. Sold the whole rig to get a 3070 laptop tho and never look back
Fantastic achievement from AMD. I expect to it be the default software solution this generation given that it’s open source and platform agnostic. A genuine gift to the industry and gamers alike 👍🏻
@@PrefoX 80%???? you realize that number you pull out of your ass is not all RTX its mostly non RTX and this FSR2.0 is opensource will run on any hardware that is powerful enough to run it.
While I'll always prefer DLSS, I can't knock AMD for its implementation of FSR in most games. It's damn fine and would be a huge boon for current generation consoles.
Weird thing to say, I prefer whichever is best. I will not always have loyalty to method A or B. Simply whichever is best and nowhere has it been written in stone that DLSS will continue to hold it's lead.
@@OGPatriot03 That's fair, but at present, DLSS is better. I think because of its dependence on hardware, it will likely remain better, too. Whatever the cause, any solution to improve visual quality and performance is good.
They were smart to add additional sharpening. There's way too much softness with TAA and it's nice to see them focusing on sharpening even if it might be over done in some areas.
I agree. Sure the skin and grass are over sharpened, but it does wonders for the side by side comparison with DLSS and TAA. I'll probably run it at more like 5 myself but it really shows the detail they're preserving with the sharpness cranked up like this.
I really think something's gone very wrong in Deathloop. Native has extremely noticeable ghosting, native is very blurry, and in some areas native even looks more aliased than either of the upscaling techniques.
Good video, as to be expected. Though i kinda feel like you analyzed in reverse. I think it's more important to see what it looks like first before discussing the performance. Also, something to keep in mind for the future, you might wanna put the lower thirrds slightly higher in comparison shots like at 10:25. Since when the video is paused, the timeline and other visual elements of the player will block the text.
That's just common youtube video editing techniques: save more important points later. And for this particular video, i'd say the performance comparison subverts common expectations, while image quality is mostly unsurprising. So it's actually important to get performance out of the way first.
Would be interesting for you to compare DLSS 2.0 to DLSS 2.4.3 that dropped with Evil Dead yesterday - too see if there is a huge difference! Great video, thanks!
I think there’s a shot. The biggest advantage is that this will run just fine on consoles, which will never be true for DLSS. That means it actually makes sense as a high-priority feature in AAA titles instead of just a halo feature for the people playing on PC.
If they do, DLSS will be easy to implement as well. In AMD's own slides describing what it takes to prepare a game for FSR 2.0, the time needed ranges from 1 month max to 3 days, with the shortest time being if it already has dlss meaning it's already set up to provide the temporal information that both techniques need (unlike FSR 1.0 which does not need temporal information). It would make sense to be true for DLSS as well--if the game supports FSR 2.0, it already provides much of what DLSS needs. At this point, the only reason I can see a game supporting one but not both techniques is if it is sponsored by AMD or nvidia (on PC at least). Which is good for everyone.
To a certain extent, yeah- but the older the GPU, the less of a return your going to see without cranking down the resolution- the card you’re using still has to be able to render at the target base resolution. So if you’re outputting at 4K using quality mode, your card still needs to be able to comfortably run the game at 1440p
Pay close attention to any game's minimum and recommended. Minimum is probably what to use to get 1080p 60fps...in most cases...Daniel Owen used a GTX 1060 and at 1080p he is getting around 50-55 fps...before FSR 2.0. For this game a recommended GPU is a RTX 2060 (or RX 5700) which he uses in this video review...the only thing missing from this review is some avg fps with 1% and .1% lows.
The reason why it doesn't work well for FSR in transparent surfaces (like particles or water) is because none of those write to the depth buffer and FSR2.0 relies on the depth to reconstruct its temporal image. Is the only downside, other than that is perfect and looks great, quite often better than native.
So based on how he explained how it works, it's FSR 1.0 but with additional stereo pass, as in the jittered frame is used to parallax effect to make stereo image. The stereo image then used to extract depth and motion flow in extension. The motion flow and depth is used to adding occlusion variables use to calculate the edges that used for the filter instead only of pixel value in FSR 1.0. Combined with temporal feature of FSR, basically you get a algorithm that could predict color, depth and motion all at once, without AI.
You got it wrong. Why would you even use stereo image for depth in a GAME ENGINE? You can literally pool perfectly accurate depth from z-buffer. This isn't real life upscaling. Jittered image is used in combination with motion vectors to generate super-sampled image. But it uses intelligent motion detection and avoids working in areas of high motion to not exhibit ghosting (which is why areas with motion appear pixelly.)
@@Navhkrinwell I kinda based my assumption from the AI SR that works from frame pixel values, forgot about z buffer honestly. but does z buffer works for text or object inside texture? I know it works for mesh edges... The jittered image works to create depth based on pixel value, works wonder for textures. The intelligent motion detection is the temporal part in FSR as in it used previous frame result in a feedback for the next frame calculation. it compares the changes in values. The high motion part is not the biggest problem, it's where the edges is represented with small amount of pixels that caused the flickering, thanks to algorithm that need to do approximation.
10:30 Yes, with 10 sharpen the skin does look rather like sandpaper, doesn't it? However, every other element looks better than Native + TAA. Notice how on the left the stitching on the gloves stands out, it looks three dimensional. On the right, it looks rather like an N64 texture, the idea of thick threading is there, however it's been blended and smeared into the polygon mesh and looks completely flat. Also, even though the sand paper sharpening is a detriment to skin, it actually makes the bumps on the leather look more realistic --- this is a problem with the skin texture itself, all other textures are improved with 10 sharpening on the left. TAA is just hiding the problems with the skin texture on the right there while crapifying all of the good textures.
Disagree. 10 is way too much sharpening. It makes everything look harsh and not blend well. At a lower level like 5 it looked much much better. But personal preferences and all that.
Great video, but I really hope developers will still continue offering DLSS alongside FSR 2.0 and not go with "easier" route with only FSR, as there are some serious motion advantages in DLSS (at least in this one game).
FSR 2 is completely different than the first, it isn't any more easy than DLSS to integrate, as a matter of fact AMD itself claim that the easiest path is with games that already support DLSS
Some minor flickering aside - Both FSR and DLSS are so fu*cking amazing! FINALLY we have technologies being pushed hard to not only show the best graphics but to also improve optimization and performance. I hope it'll mean better longevity of GPUs
This is actually an interesting development as FSR 2.0 can be added to console games if the devs choose to, in order to improve proformance during optimization and help retain more detail while meeting the desired proformance. I'm sure that games struggling with proformance issues, such as Elden Ring, as well as many games still available for last gen consoles would benefit significantly from FSR 2.0.
I remember that Alex said in one of the last podcasts that many devs are already upscaling their console versions with their own tech so it would not improve performance by much in many cases. It could improve visuals though I guess and save money and time for devs
I'm honestly pretty impressed that a human-made algorithm works this well. And really, the biggest benefit to me for FSR 2.0 is that it's open source. Of course, if you use an AMD GPU, then this is what you've been waiting for. But even if you use an NVIDIA GPU, this will be extremely useful for games that don't implement DLSS 2.0+. And really, now pretty much every game should now come with at least some good upscaling technique to increase performance while maintaining decent quality, which is awesome. Would also be nice to see some games from the last year such as RE Village implement this on consoles. Something I'm unsure of, but would love to see if possible, would be for Reshade plugins to implement this technique. Then basically any game could receive the benefits of this. I'm unsure if a developer implemented version would fair better, though. I kind of doubt it since it's more programmatic and based on previously serviced frames alone without any specifics (such as what you'd receive with deep learning algorithms). So hopefully look forward to FSR 2.0 in reshade maybe, yaay?
Doubt it becomes part of Reshade. On the bright side NVIDIA has made a developer API called Streamline so DLSS/FSR/XeSS can be added all at once instead of leaving the market so segmented.
14:40 There's a clear ghosting effect on FSR behind the radio/gadget. It just presents itself differently than with TAA. It's a super-grainy picture behind it (and that grain is in the shape of ie an antenna).
13:50 curiously on these fences FSR 2.0 does a better job, much more clean. EDIT: okay I've read your twitter post and I understand that the DLSS animation thingy goes away the moment you move again? We good. All things considered, I am glad this thing is finally out, this is a brilliant start and things can only get better from here on, AMD got back on the menu for my much awaited gaming rig upgrade. Thanks for another great video Alex.
I love that AMD is pushing open technology, but the fact that NVIDIA runs FSR even better than AMD suggests that NVIDIA is the better choice for availability of DLSS/FSR across as many games as possible.
I'm curious how much of that weird motion artifacting could be caused by the sharpening. Did you guys look at the motion artifacting examples with and without the sharpening?
10:50 Even at 0 sharpness there is a lot of flickering that isn't present when using DLSS. But of course, people are blaming Alex for using the Default level 10 sharpness set by AMD themselves because its making FSR look bad in comparison. Mindless gits need another 2kliks video where he regurgitates their own rhetoric to make them feel good.
@@AkshayKumarX and they should be, idio, he barely included any at sharpness 0. Also didn't mention deahtloop has the best implementation of DLSS, literally, most games don't even use 2.3.
As a 3080 owner i am happy to see competition and will most likely use quality mode in games that don't have DLSS support. That performance increase is well worth it.
@Bry Zeer Sounds like it took you an awfully long time to get yourself a 3080 if the 3080 Ti was selling at the same time and you could still only get that "poverty" 3080. Keep saving them checks, the 4080 is going to be outsoon -- maybe you can afford one in 2024.
Thank you Alex for shifting the focus to the important comparison points. Many other channels are mostly comparing still images which i think dont say much about the quality of a temporal upscaling solution.
I loved that at the end of the video you decided you are done with the guy throwing stones and just snapped his neck off 😂❤️ great video and very informative though 🙏
I've been waiting for the DF rundown. I think there are some aspects that look better with FSR 2.0, others better with DLSS. Kudos to AMD for making this. I would have liked however to see DLSS and FSR compared with the sharpening on FSR turned down and it turned up for DLSS.
I can't speak for FSR but DLSS sharpening is pretty much unusable if you enable HDR. The haloing becomes super apparent and is just best left disabled. However in SDR I find the sharpening slider can be useful. Hopefully it's something that can be improved upon.
Textures look like that because of sharpening. Noting to with FSR If you put sharpening slider to 0 instead of 10 it will look like DLSS However, sharpening slider only works for FSR in this game. But we have seen some games (like God of War new update) that has sharpening slider for DLSS. So the game thing can be done on DLSS
14:49 actually FSR 2.0 does show some ghosting with the device here. But it looks like it's ghosting some kind of screen space effect or the post sharpening is acting up, rather than ghosting the 3D object. In the FSR view I can see a clear trail of crunchy black pixels behind the antenna, almost looking like dithering. I wonder if maybe that's caused by how FSR does its frame accumulation, so it doesn't have enough history data to fill in the pixel gaps for those sections, resulting in a lower fidelity render.
It's the opposite of accumulation, since the antenna is occluding the objects FSR discards the history information for those pixels to avoid ghosting. It leaves some aliasing to be visible instead which the level of sharpening will make even more pronounced by increasing contrast even more.
@@MLWJ1993 this is basically exactly what I was trying to say, but phrased much better. 😂 The accumulated frames have missing data in those spots, because of the foreground object occluding the grass in the background. So FSR still tries to interpolate the missing pixel data, but does so at a much lower fidelity due to a lack of history data to pull from. Combine that with the way too aggressive sharpening filter and you end up with a crunchy pixel ghost trail haha.
Xbox for sure, but probably on PS5 too. If i'm not mistaken this july Flight Simulator should start using FSR 1.0 on Series X|S. I'd say Q1 2023 and FSR 2.0 will be on consoles too
Console titles have been using temporal image reconstruction for quite a long time now. The result varies greatly between titles, but some of them are comparable to FSR 2.0. DF discussed the power of branding for upscaling techniques, and how it's not new in console space in a recent DF weekly episode.
@@garyad8796 there are no image reconstruction technique on console that even comes close to FSR 2.0. The best among them is Insomnia's temporal injection and reconstruction which is kind of like GEN 4 TAAU.
At this point I've dug through every review of FSR 2.0 (even translated some lol), every reddit post and every TH-cam video yet I'm sure Digital Foundry will end up finding some info I had no idea about. Let's see what we got.
@@RokkieSparrow I had noticed shimmering and over sharpening in FSR. I had also noticed FSR not upscaling some alpha transparencies like the laser lines as Alex pointed out. Some new things that Alex pointed was character animation. I had not realized how much breakdown was happening there. Overall, I see there's still a decent room for improvement, but boy, for an upscaling technology that doesn't need dedicated hardware already reading 80-90% of where DLSS already is. That's seriously impressive. 1080p Performance mode is upscaling from 540p and 4k Performance mode is upscaling from 720p to give some perspective on how nuts FSR 2.0 is. It's not quite DLSS but it doesn't need to be there. I think it's amazing and I can't wait for more games to implement it.
@@RokkieSparrow Hard to say, but it's already an improvement over all temporal upscalers on console, so it's already ready to make a difference in consoles. Now if it can match DLSS, that's hard to say. DLSS uses tensor cores and leverages AI for reconstructing details, something AMD FSR by design cannot and hence it's possible that it might never top DLSS. But even reaching close without requiring additional hardware is a huge win so yay!
Ampere is faster with FSR than Navi 2 likely because of its higher memory bandwidth. It can crunch through the frame buffers of data faster. AMD's infinity cache isn't able to compensate enough for this. Thanks for the video!
i think actually it is because compute shaders run faster on nvidia, and fsr 2.0 runs on compute shaders , as it is not just ampere and high res that runs faster, like for like cards run faster too, even at lower mem bandwiths
Honestly, AMD and Nvidia should give developers the TAA part of their DLSS or FSR 2.0 and make them use it. The reason Native even loses (sometimes) to a reconstruction in many games is due to very VERY poor TAA methods being used.
Upsampling and the anti aliasing are both handled by the machine learning model in DLSS they can't be broken apart, although technically you can use DLSS as anti aliasing method like DLAA it would justr be way more expensive performance wise than standard TAA, that would work for FSR 2.0 though but it would probably still be more expensive than standard TAA.
Nice, in depth test, but I agree with others that it would have been nice to test without sharpening, both to provide and apples-to-apples comparison with the DLSS images, and to see if the sharpening adds any cost to performance.
I find it interesting that Ampere runs FSR 2.0 better than RDNA2. It would appear that Nvidia designed Ampere to be a very compute-heavy architecture, between actual CUDA cores and Tensor cores. Ampere continues to be a very forward-looking architecture. Also, love your use of the Tiberium Sun soundtrack!
At 13:19 you can see some flickering in front of the power line and the sign with white text in the FSR 2.0 test. Very small, maybe not noticeable, but not present in the native or dlss renders.
There's a already a dlss 2.39 plugin hat greatly improves the image quality of dlss 2.3 in deathloop so i wish you would've tried that but good video regardless
@@marshinz5696 I'm sorry i don't talk fanboy, but if you watched the video you can clearly still see that dlss produces better IQ. Not that it matters tho they're both fantastic and will only improve 👍🏻
@@kadupse don't bother with that one. He's the type to act high and mighty but be posting "I'm scared of zombies" under dying light videos. Her just wants attention.
Hopefully the algorithm gets some updates in the future to deal with the particle effect issues. With that it would provide a much more convicingly like-native or better than native image. The character movement ghosting issue could easily be hidden with motion blur so it's hardly much of an issue.
I feel like a lot of the problems in FSR versus DLSS can be explained by 2 things. One is that the FSR sharpening slider seems too aggressive, creating crunch and fizzle where it shouldn't be. The other is that DLSS is doing ML image analysis on the whole 3D frame, while FSR only does frame accumulation and reconstruction. It doesn't have a heuristic model to lean on when reconstructing, leading to the in motion pixel breakup and object fragmentation. It simply doesn't "understand" what objects should look like, so it's making a best effort decision about the pixels based on the frame data it does have. DLSS, with the ML model, has a heuristic understanding of the frame and objects at a shape level. It can identify and reconstruct whole shapes in an image. By contrast FSR is only capable of understanding pixel history and motion vectors. It does not "know" what shapes are and as such can sometimes cause fragmentation of shapes when it tries to match pixels together.
To my understanding, that was the case for DLSS 1.0 but it's not true for DLSS 2.x. There is no acutual ML done while running the game and it also solely relies on motion vectors color data and depth data just like FSR. The reason DLSS 2.x is better is partly because the algorithm is more mature but mostly because it can use more samples per frame than FSR to reconstruct the image thanks to the increased speed of the tensor cores.
@@TheTaxxor as far as I know both versions of DLSS use ML models computed offline. They aren't gathering training data in real time, but they are applying the heuristics of the machine learning model to the real time data being collected. DLSS 2.0 introduced a more broadly applicable model than was present in 1.0. By contrast FSR 2.0 doesn't have a neural net or machine learning model at all. It's working exclusively on data collected in real time and hand tuned heuristics, rather than machine learned heuristics.
@@TheTaxxor that said, yes the tensor cores are a significant advantage. They are almost exclusively why DLSS works, because they can accelerate the application of the neutral net. Theoretically AMD could use a neural net in FSR, but the code would probably have to run on the CPU or less efficiently on the GPU AND consume some of the raster bandwidth in the process.
I hope "mediocre but universally compatible" doesn't win out over "superior but requires specific hardware," as tends to happen. If FSR and DLSS eventually become identical in performance and quality (similar to how everyone has good anisotropic filtering now), that would be cool though.
Completely agree. I wish consoles would use nvidia instead personally. Fsr 2.0 is a big improvement but nowhere near as good as dlss. I think machine learning is just the way to go and focusing on things like fsr are the wrong direction personally. Though I'd love to be wrong
It doesn't brake sharpening but the reconstruction which create a pretty annoying pixelated effect that remind extreme sharpening, AMD traded ghosting with this artifact while the AI in DLSS is able to discard only the bad data and correctly reconstruct disoccluded pixels
14:53 This is image ruining for me, look at the left side of the device in FSR. There's a weird trail of sharpening or something that looks really odd in motion.
![[4K] Tech Focus: Volumetric Lighting - God Rays, Atmospheric/ Fog Rendering, & More Explained!](http://i.ytimg.com/vi/G0sYTrX3VHI/mqdefault.jpg)
I'm still so amazed that we can manage to produce a nicer looking image at a LOWER resolution while GAINING performance.
Absolutely blows my mind, honestly. I love tech videos like this so much!!
Yep I love that. Such a massive win. I'm assuming this will be used in console games then too, so something like Horizon they could run at 1440 FSR'd up to 4k and giving the devs some extra power left over? (Or at the very least guaranteeing a locked 60)
You got examples bc ive yet to see
@@sawsebawse2000 DLSS quality in death stranded looks better than native. Thats a well known one.
@@sawsebawse2000 in this video?
@@clownavenger0 That's becoming true for more games, to the point where it makes sense to use DLSS even just for a higher image quality in addition to performance boost.
Bottom line: If you have an RTX GPU, use DLSS. If you have anything else, you have FSR 2.0. Definitely a win for AMD and for every gamer without an RTX card.
Yeah I used to have a 1650 and I wish FSR2.0 existed back then, cause Nvidia definitely doesn't care about non-rtx GPUs
@@tazogochitashvili6514 - Nvidia doesn't really make non-RTX GPUs any more.
I basically own only Nvidia GPU's but you gotta love AMDs efforts they literally transformed by 1080 with fsr and now even more with fsr 2.0!!
@@tazogochitashvili6514 lmao yes Nvidia doesn't give a fuck about its customers amd doesn't care too much either since both only look for money Ofc but AMD takes the role of the good cop and nvidia is the ahole cop f'ing over everyone that hasn't bought a product of theirs in the last 2 years
Edit: and also those that buy their products regularly, so basically everyone
FSR benefits everyone. There will likely be developers that choose to invest into implementing FSR, but not DLSS since it will benefit a smaller proportion of their player base. This is an absolute win for gamers, thank you AMD (BTW I only have nvidia cards)
This is a big improvement and awesome for those who don't have a DLSS capable card.
I have a 3070 and I'm going to use FSR 2.0 because it's clearer in motion.
Big improvement? We rarely play games in still image. This has artefacts in motion which is a big no no. This will make the overall image look way shimmery in motion much like HFW performance mode.
@@jal.ajeera Reading comprehension. It's a big improvement from FSR 1.0.
@@letsgoiowa Uh, no it's not??
@@Downloadfreak It's not. Did you see the breaking apart of image in motion at performance mode reconstruction? Shimmering extravaganza! How is that an improvement?
Audio and visual comprehension much? Lol.
FSR 2.0 is gonna do wonders for Series X/S and PS5
Steam Deck !
Yep specially the less capable series S
@@jacobartzavia6096 Exactly! Also the lower memory bandwidth on Series S will do better with an FSR-upscaled image from 1080p/720p rather than a native 1440p in games.
What's the betting pool on Deathloop getting an FSR2.0 update on PS5?
This will be big news for Minecraft’s unreleased DXR path tracing
I love that DLSS gets rid of that awful shimmering in so many games and it's nice to see FSR should at least improve things in many cases too. For some reason when I see objects like fences doing that in games it has always bugged me.
Aliasing - the oldest problem in 3D gaming, perhaps lol.
@@hi_tech_reptilez I would say transparency is a much bigger challenge. Even with raytracing and 3090 games need tons of hacks to get away with faking it.
@@hi_tech_reptilez Personally I don't mind Aliasing as much as motion blur, with the lack of progress on non temporal AA solutions ghosting and blur has become the biggest problem with modern gaming for me as often there is no way to turn TAA off in the settings menu.
dll wins easy. shame on amd
Dynamic link library doesn't have much to do with AMD
The default sharpness of 10 is definitely way too much in my experience, and lowering to 5 or 0 fixes alot sharpness artifacts
Yeah this eliminates a lot of the weird aliasing type artifacts
They probably did that intentionally so that in side by side comparisons with DLSS the details on geometry etc look sharper. Hopefully they tune the default down once they're not just trying to prove themselves.
Yeah I was thinking that, 10 seems a bit too high which could be causing a few of the other problems, it's a shame Ales didn't show some of those areas with the sharpness set to 7, 5, and 3 to see the results.
It drove me nuts the whole video. Basically everything he pointed out as an issue was an issue with the sharpness being up too much.
Yeah not sure what the devs were thinking, the image is obviously oversharpened at the default setting
As someone who works with FSR 1.0's code a lot, and this has the same sharpness pass I have always found that 6-9 was the sweet spot, never used a game where I had to go lower or higher (unless the games TAA was a super blurry mess however this replaces the games AA) so 10 should realistically never be required. If you can't be bothered to test for the best value just set it to 8 and forget it though, that's my tip.
Seeing as DLSS will never come to current gen consoles, this is a huge step up for them compared to native rendering/gen 1 upscaling. Of course, there’s also UE5’s TSR on consoles, but that will be restricted to UE5 games. Would love to see a comparison video of MS flight sim after FSR 2.0 is implemented.
Why not ?
@@puffy3146 DLSS is exclusive to Nvidia cards, specifically gpus’s with tensor cores/machine learning (RTX 2060 and above). Current gen consoles use AMD hardware that is not compatible with DLSS.
@@williamcrms i see, so will we at least get the fsr2.0 this gen ? Series n ps5 or do you think later on in the gen
@@puffy3146 I think we’ll start seeing it used more on console beginning in 2023 but it’s up to individual studios to implement it so have to wait and see.
@@williamcrms thanks bud !
Amazing technology. How did someone come up with this stuff, it’s so bloody complicated. Bravo to AMD for making it applicable across competitor GPU’s.
It is amazing. Essentially giving us a small hardware upgrade in the form of software. For free.
bravo to AMD? lol without Nvidia there would be no FSR at all, and DLSS' tech is at least 2 years further
@@PrefoX without AMD there would be not Upscaling for GTX-Cards because Nvidia doesn't care about GTX-Users
This is the natural evolution of TAA, where you attempt to use the additional information from previous frames to improve the output detail. This is "just" retaining more useful data so you can usually increase the output resolution as well as detail. The "just" is that this is actually horribly complicated to do without artifacts like "slime trails", strobing, or shimmer - like most things, the idea is a lot easier than the work!
@@SimonBuchanNz Indeed! I still remember how early 2015/2016 TAA had a LOT of ghosting.
The most competent analysis I've read since the FSR 2.0 release (only yesterday)! Such a good job, in such a short amount of time. Anyway the gap between FSR 1.0 and 2.0 is huge and it will grow for sure with the next updates!
Hardware Unboxed is also worth of watching methinks in that regard.
@@GameslordXY yep, the HU review sounds more interesting than this.
This is pretty interesting. AMD have made some nice advancements with FSR, hopefully the ghosting can be patched out. Either way, it's nice to see more competition in this area of rendering.
by it's very nature temporal based AA will always by its nature create ghosting. It's why MSAA always looks nicer as its mid process and actually takes 3D information into account instead of just the depth buffer.
The actual pixelation you can sometimes see seems to also be partially caused by AMD's implementation to avoid ghosting (i.e. discarding the history buffer for occluded pixels, something you can see happening with the sidemovement with the antenna).
I have a feeling FSR 2.0 will be like DLSS is, where the quality of it's use is on a per-game basis, some do it better than others, so yeah, like Alex said, we need to see more games with FSR 2.0 but so far, the results look really promising and it's likely only going to get better with future revisions of FSR.
Yes, I'm familiar with different methods of AA and how they're implemented. However, as was noted in the video, this is just ONE implementation of FSR2.0, in a single game. Time and further development will tell how this plays out in in other games.
@@Draggobuttboi ,
There ARE solutions to fix the ghosting that TAA tends to cause.
You can analyze the surrounding pixels and provide an upper limit on brightness for the aliased part. Similarly there are ways to use the motion vector values to apply motion compensation.
So it's NOT an unfixable problem.
I'm not sure how much VRR and/or Dynamic Resolution affects these workarounds.
Alex is an absolute beast. Tremendously informative and technical analysis. Congratulations.
Extremely interesting. Excited to see this on Microsoft Flight Simulator and Forspoken. Was extremely lucky to find a 6800xt at MSRP in 2020, but I have been missing DLSS. Happy there's competition now and that it's open sourced.
Super informative dive into the various stress tests for the reconstruction techniques. FSR 2.0 looks exciting. Not as robust as DLSS, but the broader hardware applicability and performance savings in quality mode are great to see. I'm hoping for some future integration into XSX and PS5 to hit 4k60 from either 1080p or 1440p internal resolutions in games with engines that would otherwise not have such a good reconstruction techniques available to them
Wow, this is an absolute game changer. Especially in the console space. With this- the PS5 & Xbox Series X can now conceivably run every title at 4K/60fps and with an image quality that is arguably better than native. On the PC side it’s great that this also works on any GPU. Cant wait to test it out myself
Game developers better use it, I swear to god...
@@LouSassol69er Sadly most don’t. I’ve only had a DLSS capable GPU for around 6 months and the tech is incredible. It’s astounding to me that this isn’t implemented in every single game. With FSR 2.0 there is no longer an excuse- because you don’t even need an RTX card or dedicated tensor cores, it just works on everything.
A lot of console games were probably already doing something like this internally. But many temporal upscaling solutions are clearly more rudimentary than this, it's good to have an open source solution that everyone can use, and presumably submit patches back to.
@@liamness This is light years ahead of what has been implemented on the console side before.
It’s another tool in your bag of tricks.
I keep hearing "better than native". What an exciting time to be alive.
Since it's a relatively new technology I imagine there's still a lot for game devs to learn when it comes to using it to it's max
Exactly, in five years this would be standard and that's awesome
@@marcomarco86tecno probably sooner than that.
The lateral test zoomed in on the device was the most interesting. TAA seems to be ghosting a lot as it is weighting samples too evenly such that stale data is still present and contributing strongly to pixel color, DLSS seems to be more cleverly weighting its samples by giving stale samples much lower influence on pixel color but if a sample is that stale it should probably just be evicted, its still ghosting just not as badly as TAA, meanwhile FSR is taking an aggressive stance on ghosting by evicting stale samples with prejudice this removes the trail of the device, BUT that means the recently disoccluded grass is now noticeably under sampled making it noisy and pixelated.
DLSS 2 is now almost 3 years old, I wonder if they are cooking something new in the oven...
As a gtx 1070 scrub, this is great news. Can’t wait for fsr 2.0 to hit dying light 2 as I found the fsr 1.0 implementation in that game to be a very poor solution
That game is just kinda fugly no matter what setup you have.
Finally. It bugged the crap out of me that people kept comparing FSR to DLSS when they weren't trying to do the same thing at all.
Now that comparison actually makes sense.
Ooh been looking forward to DF looking at this, so much confusion amongst other news sources about what FSR is and what it does.
Disgusting creature
Every time Alex uses music from Tiberian Sun in a video, it just reminds me how nice it would be to have a remaster.
Thank you I knew I recognized it from somewhere
Sadly with EA it sent Westwood and it's games to the grave like many others
I'm just waiting for all this tech to be used as normal anti aliasing. I'm sick of the super soft look of TAA and ghosting in motion.
Because these upscalers are designed to be as crisp and 'sharp' as possible, I'm guessing they'd produce significantly better results than TAA if they could use native resolution.
Here to see john's thread arguments =]
@@aweigh1010 it can certainly be implemented pretty well, a lot of frostbite games look quite good with it, but you only need to play halo infinite or almost any recent ue4 game to realise it can have serious issues.
Anyway, aside from whether you prefer it over older methods, it would be ignorant to completely ignore its flaws. Flaws that dlss and fsr 2.0 do a decent job of avoiding, even with their lower resolutions.
Doom 2016/Doom eternal use TSSAA which is the best form of TAA I've seen implemented with hardly noticeable ghosting
fsr 1.0 is currently on switch
@@FarmedRice haven't personally played doom, but I don't doubt that great TAA Implementations can exist, but if there's going to be a push to include dlss or fsr in newer games, it would be nice if it came along with a native resolution option too. Just for all those games out there with worse TAA than doom
1.05 ms added for a chip without dedicated silicon is simply awesome. I wish that we had 2022 image processing technologies in 2003. Would have made the transition to flat panels so much less annoying back then.
I’m not very tech savvy so I didn’t understand the added ms. What does that mean in terms of FPS? So if a game’s running at 60fps and we turn on FSR 4K Quality mode, what can we expect the FPS to be abouts?
@@TheHulksMistress Every frame takes time to render, this is often measured in milliseconds. I think 16.6 milliseconds per frame is needed to hit 60 fps. The added milliseconds is how much extra time is taken to render the effect. For example FSR 2.0 may take an extra 2ms to render, however, lowering the resolution by so much saves more time than it adds. Therefore you get more frames, so if a game is running at 60 at native 4k, and you turn on FSR 2.0 quality, you almost certainly will be above 80fps or higher.
That might actually get better with driver and FSR optimizations but for the start, it's already in a good state.
It will be interesting to see with driver and FSR optimizations if that will translate to all FSR 2.0 games or if it's on a game per game basis.
@@Thyispro great, thanks for the example. I love that. I’ve watched enough Digital Foundry videos to know what devs can do with an extra 20fps worth of power to play with
Speaking of LCD's..I remember my dual 19" Princeton Graphics CRTs' back in the day XD...The model was Princeton Ultra 97. 1600 x 1200 resolution on both monitors...of course at the time I had a GT 9500 back then.
I wouldn't be surprised if FSR 2.0 started showing up as games as the default setting, in PC and consoles with it looking as good as it does and improving performance so much at the same time.
Yeah in cyberpunk dlss is default
@4K 60FPS for sure man, but god's know how long
Maybe longer term but I doubt that will happen for a few years, but from a developer's point of view, having that on as default would lower the hardware requirements needed, that could mean more game sales.
@@paul1979uk2000 less effort to put in making it optimised enough to run
Sony has to add it to their sdk so God only knows how long it's gonna take those useless fucks. We got vrr 3 yrs later have yet to get texture filtering for every game like Xbox does and 120hz. So yeah good luck with that.
Always good to see your extremely thorough testing, finding issues that no other outlets seem to notice. It's certainly a huge improvement over FSR 1.0 that will be very welcome for many, but I'll definitely continue using DLSS when I can.
One thing you didn't touch on that's been talked about quite a bit though - anti-aliasing quality on slanted lines. Many people seem to be under the impression that FSR 2.0 actually handles this better, though it seems like about a tie to me. I'd be interested to hear your take on that, Alex.
I was watching a second time and I noticed that he also does not take segments where DLSS suffers from ghosting, because generally DLSS suffers from ghosting, and that happens in this game as well. One of the big surprises in all the video comparisons so far was that FSR 2.0 handles ghosting way better than DLSS, yet here in this video Alex almost succeeded in making it look the other way around 14:41
@@ctrlectrld Well different test cases will inevitably provide different results. It is odd that he didn't seem to produce the DLSS weapon sight ghosting that other reviews have noted (but also didn't say anything about the severe grainy trail on the grass on FSR 2.0 in the ghosting test.) I do wonder if some of these issues might be caused by the sharpening pass though and will look better with it turned down a bit.
@@nimbulan2020 Yep, really weird. He wrote on twitter that sharpening is the reason for the grainy trail. Thing is, people noticed the grainy trail, because it's very noticeable, and that's maybe why he didn't say anything about it, cause he didn't need to. But again, this way I am not sure he's presenting things fairly. As things stand, if we only had this video as reference, we would come to the conclusion that FSR 2.0 handles ghosting worse than DLSS, which is quite unfortunate.
@@ctrlectrld Well is it any less fair than other reviews not even mentioning the very noticeable issues that FSR 2.0 has? Realistically, reviewers have a limited about of time to test things, and they're not always going to catch every issue or be able to test every configuration. Alex stuck with the default sharpening setting since that's the one people will most likely use. It may look better in other games, we'll have to see when the time comes.
@@nimbulan2020 It seems to me that other relevant reviews, like the one by Hardware Unboxed, show the good and the bad.
But in my previous comment I was referring to the ghosting issue specifically; in this case I feel that the topic is not being addressed fairly since ghosting has always been a main issue in the temporal domain, so much that DLSS is still struggling with it to this day. AMD with FSR 2.0 wanted to tackle specifically this issue and came up with a clever solution utilizing the depth buffer to reduce ghosting artifacts. It does a good job, at least better than DLSS, and it's one of the main perks of FSR 2.0. You can't just leave this whole thing out from a specialized in depth tech review, and on top of that even pick a segment where the better solution looks worse. It's not a fair and complete representation in my opinion.
This is really welcome for someone like myself, I have a GTX 1070 but am holding out for a 30/40 series that I really can't afford right now.
I feel like 1070 and 1080 owners are the biggest benefactors of this news.
@@DanKaschel $700 1080ti 2017 is one of the best purchase I made. This card is still decent today, especially if fsr 2.0 is coming.
Sold the whole rig to get a 3070 laptop tho and never look back
Fantastic achievement from AMD. I expect to it be the default software solution this generation given that it’s open source and platform agnostic. A genuine gift to the industry and gamers alike 👍🏻
80% of the gamers use a Nvidia card, why would FSR2 be the default then? because the work to implement is the same as DLSS.
@@PrefoX consoles basically
@@PrefoX 80% of the gamers? Where's that number from? There are many AMD-based consoles out there
@@PrefoX 'because the work to implement is same as DLSS' - source??
@@PrefoX 80%???? you realize that number you pull out of your ass is not all RTX its mostly non RTX and this FSR2.0 is opensource will run on any hardware that is powerful enough to run it.
While I'll always prefer DLSS, I can't knock AMD for its implementation of FSR in most games. It's damn fine and would be a huge boon for current generation consoles.
Weird thing to say, I prefer whichever is best. I will not always have loyalty to method A or B.
Simply whichever is best and nowhere has it been written in stone that DLSS will continue to hold it's lead.
@@OGPatriot03 That's fair, but at present, DLSS is better. I think because of its dependence on hardware, it will likely remain better, too. Whatever the cause, any solution to improve visual quality and performance is good.
cant' use both methods at the same time? fsr 2.0 + dlss 2.0 ? someone know if is posible in the future?
@@SuperMatias24 I don't think you'd really want to just both at the same time lol
Always prefer DLSS? Not fanboi words at all.. 🙄
As a GTX 1660Ti owner, I desperately need this for Cyberpunk! 😫
By the next update for that game, it should be there.
Wow, alex. Amazing work as ever. 15:49 especially, that is a very good comparison
FSR 2.0 is really impressive no matter if you're team red or green with a few more updates watch out dlss
Team green isn't going to be using FSR 2.0 if they have the option of DLSS.
@@icy1007 Not if you don't have an RTX card.
They were smart to add additional sharpening. There's way too much softness with TAA and it's nice to see them focusing on sharpening even if it might be over done in some areas.
I would lower that setting quite a bit. maybe try 3 or 5
I agree. Sure the skin and grass are over sharpened, but it does wonders for the side by side comparison with DLSS and TAA. I'll probably run it at more like 5 myself but it really shows the detail they're preserving with the sharpness cranked up like this.
I really think something's gone very wrong in Deathloop.
Native has extremely noticeable ghosting, native is very blurry, and in some areas native even looks more aliased than either of the upscaling techniques.
Good video, as to be expected. Though i kinda feel like you analyzed in reverse. I think it's more important to see what it looks like first before discussing the performance. Also, something to keep in mind for the future, you might wanna put the lower thirrds slightly higher in comparison shots like at 10:25. Since when the video is paused, the timeline and other visual elements of the player will block the text.
That's just common youtube video editing techniques: save more important points later. And for this particular video, i'd say the performance comparison subverts common expectations, while image quality is mostly unsurprising. So it's actually important to get performance out of the way first.
@@garyad8796 that definitely makes sens. I forgot watchtime is so important here :D
I actually preferred how it's done in the video. If it's not performant then i don't care what it looks like.
But I can understand your argument.
@@pixelslaughter3492 Sounds fair enough.
Would be interesting for you to compare DLSS 2.0 to DLSS 2.4.3 that dropped with Evil Dead yesterday - too see if there is a huge difference! Great video, thanks!
Great work Alex!
With your analysis, you have put to shame all other comparisons of this technology that I have seen on the web.
This is a good alternative to DLSS, specially for those who don't have a RTX GPU. DLSS and/or FSR 2.0 need to become a standard on consoles too
DLSS isn't possible on current gen consoles.
FSR of course totally is, and they were obviously designed with this in mind.
@@tzuyd If its cost at 4k is already 2 full ms (above FSR1) on an rx6800 I wonder how will it run on a console
i have a rtx 2070 super but god i hope fsr 2.0 actually pushes developers to implement it in every game
I think there’s a shot. The biggest advantage is that this will run just fine on consoles, which will never be true for DLSS. That means it actually makes sense as a high-priority feature in AAA titles instead of just a halo feature for the people playing on PC.
If they do, DLSS will be easy to implement as well. In AMD's own slides describing what it takes to prepare a game for FSR 2.0, the time needed ranges from 1 month max to 3 days, with the shortest time being if it already has dlss meaning it's already set up to provide the temporal information that both techniques need (unlike FSR 1.0 which does not need temporal information). It would make sense to be true for DLSS as well--if the game supports FSR 2.0, it already provides much of what DLSS needs.
At this point, the only reason I can see a game supporting one but not both techniques is if it is sponsored by AMD or nvidia (on PC at least). Which is good for everyone.
It's almost hard for me to believe that GPU vendors create this amazing tech. It essentially allows you to postpone upgrading your existing GPU.
To a certain extent, yeah- but the older the GPU, the less of a return your going to see without cranking down the resolution- the card you’re using still has to be able to render at the target base resolution. So if you’re outputting at 4K using quality mode, your card still needs to be able to comfortably run the game at 1440p
@@rickrodriguez6244 Agreed! But if you buy a new GPU right now you can be more confident about it retaining its value for longer.
@@EmblemParade Definitely
What graphics developers did in the last 10 years is voodoo.
Pay close attention to any game's minimum and recommended. Minimum is probably what to use to get 1080p 60fps...in most cases...Daniel Owen used a GTX 1060 and at 1080p he is getting around 50-55 fps...before FSR 2.0. For this game a recommended GPU is a RTX 2060 (or RX 5700) which he uses in this video review...the only thing missing from this review is some avg fps with 1% and .1% lows.
Now that DF has weighed in, I can form an opinion. Thank you gentlemen!
The reason why it doesn't work well for FSR in transparent surfaces (like particles or water) is because none of those write to the depth buffer and FSR2.0 relies on the depth to reconstruct its temporal image. Is the only downside, other than that is perfect and looks great, quite often better than native.
So based on how he explained how it works, it's FSR 1.0 but with additional stereo pass, as in the jittered frame is used to parallax effect to make stereo image. The stereo image then used to extract depth and motion flow in extension. The motion flow and depth is used to adding occlusion variables use to calculate the edges that used for the filter instead only of pixel value in FSR 1.0. Combined with temporal feature of FSR, basically you get a algorithm that could predict color, depth and motion all at once, without AI.
You got it wrong. Why would you even use stereo image for depth in a GAME ENGINE? You can literally pool perfectly accurate depth from z-buffer. This isn't real life upscaling.
Jittered image is used in combination with motion vectors to generate super-sampled image. But it uses intelligent motion detection and avoids working in areas of high motion to not exhibit ghosting (which is why areas with motion appear pixelly.)
@@Navhkrinwell I kinda based my assumption from the AI SR that works from frame pixel values, forgot about z buffer honestly. but does z buffer works for text or object inside texture? I know it works for mesh edges... The jittered image works to create depth based on pixel value, works wonder for textures. The intelligent motion detection is the temporal part in FSR as in it used previous frame result in a feedback for the next frame calculation. it compares the changes in values. The high motion part is not the biggest problem, it's where the edges is represented with small amount of pixels that caused the flickering, thanks to algorithm that need to do approximation.
no
Looks like playable framerates are back on the menu, boys!
Sub 60 FPS are not for playing!
@@AIIEYESONME My 1070 can finally run cyberpunk 1080p medium with 60+ fps all of the time, like a champ! Ty amd for being a bro for pascal users
Loving the use of some classic command and conquer music in the background!
10:30
Yes, with 10 sharpen the skin does look rather like sandpaper, doesn't it?
However, every other element looks better than Native + TAA.
Notice how on the left the stitching on the gloves stands out, it looks three dimensional.
On the right, it looks rather like an N64 texture, the idea of thick threading is there, however it's been blended and smeared into the polygon mesh and looks completely flat.
Also, even though the sand paper sharpening is a detriment to skin, it actually makes the bumps on the leather look more realistic --- this is a problem with the skin texture itself, all other textures are improved with 10 sharpening on the left.
TAA is just hiding the problems with the skin texture on the right there while crapifying all of the good textures.
You're totally right, the FSR image is overall better than the native one, it just has some small patches that look worse.
Disagree. 10 is way too much sharpening.
It makes everything look harsh and not blend well. At a lower level like 5 it looked much much better.
But personal preferences and all that.
love the old df music playing in the background. takes me back to 2013 watching comparisons of COD Ghost!
Great video, but I really hope developers will still continue offering DLSS alongside FSR 2.0 and not go with "easier" route with only FSR, as there are some serious motion advantages in DLSS (at least in this one game).
FSR 2 is completely different than the first, it isn't any more easy than DLSS to integrate, as a matter of fact AMD itself claim that the easiest path is with games that already support DLSS
Some minor flickering aside - Both FSR and DLSS are so fu*cking amazing! FINALLY we have technologies being pushed hard to not only show the best graphics but to also improve optimization and performance.
I hope it'll mean better longevity of GPUs
This is actually an interesting development as FSR 2.0 can be added to console games if the devs choose to, in order to improve proformance during optimization and help retain more detail while meeting the desired proformance. I'm sure that games struggling with proformance issues, such as Elden Ring, as well as many games still available for last gen consoles would benefit significantly from FSR 2.0.
Don't get your hopes up for Elden Ring.
The devs are making fundemental mistakes that affect performance.
I remember that Alex said in one of the last podcasts that many devs are already upscaling their console versions with their own tech so it would not improve performance by much in many cases. It could improve visuals though I guess and save money and time for devs
@@Dionyzos hope many games will get an upgrade, so i can enjoy as example ac origins in 60fps as soon as possible on series x.
Finally some tests that show how these techniques behave in motion!
I'm honestly pretty impressed that a human-made algorithm works this well. And really, the biggest benefit to me for FSR 2.0 is that it's open source. Of course, if you use an AMD GPU, then this is what you've been waiting for. But even if you use an NVIDIA GPU, this will be extremely useful for games that don't implement DLSS 2.0+. And really, now pretty much every game should now come with at least some good upscaling technique to increase performance while maintaining decent quality, which is awesome. Would also be nice to see some games from the last year such as RE Village implement this on consoles. Something I'm unsure of, but would love to see if possible, would be for Reshade plugins to implement this technique. Then basically any game could receive the benefits of this. I'm unsure if a developer implemented version would fair better, though. I kind of doubt it since it's more programmatic and based on previously serviced frames alone without any specifics (such as what you'd receive with deep learning algorithms). So hopefully look forward to FSR 2.0 in reshade maybe, yaay?
Doubt it becomes part of Reshade. On the bright side NVIDIA has made a developer API called Streamline so DLSS/FSR/XeSS can be added all at once instead of leaving the market so segmented.
FSR 2.0 cannot be implemented in Reshade. This would be too good to be true
This video showed the differences much better than any other channel did. Well done as always
14:40 There's a clear ghosting effect on FSR behind the radio/gadget. It just presents itself differently than with TAA. It's a super-grainy picture behind it (and that grain is in the shape of ie an antenna).
I wish more of the comparisons showed FSR Quality mode (unless I missed it) a lot only showed performance mode...
Watched about five reviews of FSR 2.0 and this is by far the best!
13:50 curiously on these fences FSR 2.0 does a better job, much more clean. EDIT: okay I've read your twitter post and I understand that the DLSS animation thingy goes away the moment you move again? We good.
All things considered, I am glad this thing is finally out, this is a brilliant start and things can only get better from here on, AMD got back on the menu for my much awaited gaming rig upgrade.
Thanks for another great video Alex.
I love that AMD is pushing open technology, but the fact that NVIDIA runs FSR even better than AMD suggests that NVIDIA is the better choice for availability of DLSS/FSR across as many games as possible.
I'm curious how much of that weird motion artifacting could be caused by the sharpening. Did you guys look at the motion artifacting examples with and without the sharpening?
0
They can only test so many variables. Makes sense to prioritize 10 since that is the default.
10:50 Even at 0 sharpness there is a lot of flickering that isn't present when using DLSS.
But of course, people are blaming Alex for using the Default level 10 sharpness set by AMD themselves because its making FSR look bad in comparison.
Mindless gits need another 2kliks video where he regurgitates their own rhetoric to make them feel good.
@@AkshayKumarX and they should be, idio, he barely included any at sharpness 0. Also didn't mention deahtloop has the best implementation of DLSS, literally, most games don't even use 2.3.
jeez what a workload for this vid lol appreciate yall!!
Would be very interesting to see FSR 2.0 up against Epic’s TSR, feels like they behave very similarly in motion
I would hope FSR2 is at least 20-30 percent faster, because DLSS is 30-80 percent faster in every case vs TSR UE5.
Saw lots of FSR 2.0 videos in my feed but I’ll always wait for Digital Foundry when it comes to stuff like this.
As a 3080 owner i am happy to see competition and will most likely use quality mode in games that don't have DLSS support. That performance increase is well worth it.
@Bry Zeer lol "weak" 3080. This is like having a normal yacht instead of the Jeff Bezo yacht. It's still better than 99% of what's out there.
@Bry Zeer damn bro you bought the Ti for that 5% extra performance, real chad move.
@@birb_person Jensen's coffers are pleased.
@Bry Zeer 3080 Ti is still only a ~5% gain lol
@Bry Zeer Sounds like it took you an awfully long time to get yourself a 3080 if the 3080 Ti was selling at the same time and you could still only get that "poverty" 3080. Keep saving them checks, the 4080 is going to be outsoon -- maybe you can afford one in 2024.
Thank you Alex for shifting the focus to the important comparison points. Many other channels are mostly comparing still images which i think dont say much about the quality of a temporal upscaling solution.
Hot Wheels Unleashed supported AMD FSR 1.0 and it looked amazing! Now with 2.0 released I can't wait for other games to implement it.
I loved that at the end of the video you decided you are done with the guy throwing stones and just snapped his neck off 😂❤️ great video and very informative though 🙏
I've been waiting for the DF rundown. I think there are some aspects that look better with FSR 2.0, others better with DLSS. Kudos to AMD for making this. I would have liked however to see DLSS and FSR compared with the sharpening on FSR turned down and it turned up for DLSS.
I can't speak for FSR but DLSS sharpening is pretty much unusable if you enable HDR. The haloing becomes super apparent and is just best left disabled. However in SDR I find the sharpening slider can be useful. Hopefully it's something that can be improved upon.
Backround Music is back :D With Ray Tracing Alex :D
Solid textures for me look best on FSR 2.0, but DLSS clearly has an advantage on anything that's only a few pixels thin like fences and wires.
Textures look like that because of sharpening. Noting to with FSR
If you put sharpening slider to 0 instead of 10 it will look like DLSS
However, sharpening slider only works for FSR in this game. But we have seen some games (like God of War new update) that has sharpening slider for DLSS. So the game thing can be done on DLSS
The "und auf Wiedersehen" at the end killed me. Funny stuff.
14:49 actually FSR 2.0 does show some ghosting with the device here. But it looks like it's ghosting some kind of screen space effect or the post sharpening is acting up, rather than ghosting the 3D object. In the FSR view I can see a clear trail of crunchy black pixels behind the antenna, almost looking like dithering.
I wonder if maybe that's caused by how FSR does its frame accumulation, so it doesn't have enough history data to fill in the pixel gaps for those sections, resulting in a lower fidelity render.
That looks more like oversharpening artifacts that maybe amplifying a slight inverse ghosting trail?
It's the opposite of accumulation, since the antenna is occluding the objects FSR discards the history information for those pixels to avoid ghosting. It leaves some aliasing to be visible instead which the level of sharpening will make even more pronounced by increasing contrast even more.
@@MLWJ1993 this is basically exactly what I was trying to say, but phrased much better. 😂
The accumulated frames have missing data in those spots, because of the foreground object occluding the grass in the background. So FSR still tries to interpolate the missing pixel data, but does so at a much lower fidelity due to a lack of history data to pull from. Combine that with the way too aggressive sharpening filter and you end up with a crunchy pixel ghost trail haha.
So you're telling me it's a technique similar to image stacking but for cinematography. Awesome 👌
Great content. A question: will it come to PS5 and XSX/S?
Xbox for sure, but probably on PS5 too. If i'm not mistaken this july Flight Simulator should start using FSR 1.0 on Series X|S. I'd say Q1 2023 and FSR 2.0 will be on consoles too
Console titles have been using temporal image reconstruction for quite a long time now. The result varies greatly between titles, but some of them are comparable to FSR 2.0. DF discussed the power of branding for upscaling techniques, and how it's not new in console space in a recent DF weekly episode.
@4K 60FPS As i said before, Xbox has confirmed it, we don't know about PS5 yet (it could, but until Sony says anything we don't know)
@@garyad8796 there are no image reconstruction technique on console that even comes close to FSR 2.0. The best among them is Insomnia's temporal injection and reconstruction which is kind of like GEN 4 TAAU.
This is the first video in which I've caught Alex using a contraction.
At this point I've dug through every review of FSR 2.0 (even translated some lol), every reddit post and every TH-cam video yet I'm sure Digital Foundry will end up finding some info I had no idea about. Let's see what we got.
Whachu think?
@@RokkieSparrow I had noticed shimmering and over sharpening in FSR. I had also noticed FSR not upscaling some alpha transparencies like the laser lines as Alex pointed out.
Some new things that Alex pointed was character animation. I had not realized how much breakdown was happening there.
Overall, I see there's still a decent room for improvement, but boy, for an upscaling technology that doesn't need dedicated hardware already reading 80-90% of where DLSS already is. That's seriously impressive. 1080p Performance mode is upscaling from 540p and 4k Performance mode is upscaling from 720p to give some perspective on how nuts FSR 2.0 is.
It's not quite DLSS but it doesn't need to be there. I think it's amazing and I can't wait for more games to implement it.
@@youtindia do you think it’ll be as good or anywhere near as good or even available on console?
@@RokkieSparrow Hard to say, but it's already an improvement over all temporal upscalers on console, so it's already ready to make a difference in consoles.
Now if it can match DLSS, that's hard to say. DLSS uses tensor cores and leverages AI for reconstructing details, something AMD FSR by design cannot and hence it's possible that it might never top DLSS.
But even reaching close without requiring additional hardware is a huge win so yay!
@@youtindia yayy
Loving the Tiberian Sun soundtrack.
Ampere is faster with FSR than Navi 2 likely because of its higher memory bandwidth. It can crunch through the frame buffers of data faster. AMD's infinity cache isn't able to compensate enough for this. Thanks for the video!
i think actually it is because compute shaders run faster on nvidia, and fsr 2.0 runs on compute shaders , as it is not just ampere and high res that runs faster, like for like cards run faster too, even at lower mem bandwiths
Nice to see such great results without the need for closed off hardware that most people don't have access to.
Honestly, AMD and Nvidia should give developers the TAA part of their DLSS or FSR 2.0 and make them use it.
The reason Native even loses (sometimes) to a reconstruction in many games is due to very VERY poor TAA methods being used.
Upsampling and the anti aliasing are both handled by the machine learning model in DLSS they can't be broken apart, although technically you can use DLSS as anti aliasing method like DLAA it would justr be way more expensive performance wise than standard TAA, that would work for FSR 2.0 though but it would probably still be more expensive than standard TAA.
Nice, in depth test, but I agree with others that it would have been nice to test without sharpening, both to provide and apples-to-apples comparison with the DLSS images, and to see if the sharpening adds any cost to performance.
I find it interesting that Ampere runs FSR 2.0 better than RDNA2. It would appear that Nvidia designed Ampere to be a very compute-heavy architecture, between actual CUDA cores and Tensor cores. Ampere continues to be a very forward-looking architecture.
Also, love your use of the Tiberium Sun soundtrack!
Same for Turing and NVIDIA received a lot of undeserved criticism for it, time proven them right
Thanks for this in depth analysis.
Gonna have to play some C&C TS now... X)
At 13:19 you can see some flickering in front of the power line and the sign with white text in the FSR 2.0 test. Very small, maybe not noticeable, but not present in the native or dlss renders.
It is actually present in Native as well, just a lot less noticeable.
An outstanding analysis by Alex, it shows great experience and attention to detail. Team DF always delivers.
There's a already a dlss 2.39 plugin hat greatly improves the image quality of dlss 2.3 in deathloop so i wish you would've tried that but good video regardless
Where can you get it?
AMD won. Get over it. You paid for nothing.
@@marshinz5696Did you even watch the video?
@@marshinz5696 I'm sorry i don't talk fanboy, but if you watched the video you can clearly still see that dlss produces better IQ. Not that it matters tho they're both fantastic and will only improve 👍🏻
I doubt it. The human eye cant see more than DLSS 2.2
All these images reconstruction technique really are a "God" sent for lower end cards and consoles.
I’d like to see this in Forbidden West on PS5. I don’t like the checkerboard rendering of the current 60fps mode, this would be much cleaner
Guerilla don’t care lol
@@JamieRobert_ Uh. Ok.
@@JamieRobert_ they have an update improving the shimmering. hopefully they support FSR 2.0
@@LeafyTheLeafBoy they’ve been saying that for months and nothing has changed they’re lazy
@@JamieRobert_ Ignorance and entitlement; name a more iconic duo.
Amazing tech as a general purpose solution. Can't wait for this to be implemented in more titles.
i found out nvidia architecture tend to be faster in pure shader computing so if fsr 2.0 is shader based it could be why nvidia tend to be faster .
i agree i think this is why, compute shaders more efficient on Nvidia. They have focused on gpu compute a lot for a while now in cuda
Props for the Command and Conquer music!
FSR was just added to Deep Rock Galactic in the latest patch.
A game changer.
PC or console? Or both?
@@ThaGeneralify PC
@@kadupse don't bother with that one. He's the type to act high and mighty but be posting "I'm scared of zombies" under dying light videos. Her just wants attention.
Dude you are a graphics monk. Thanks for the breakdown!
DLSS is better but FSR 2.0 is impressive and open
Hopefully the algorithm gets some updates in the future to deal with the particle effect issues. With that it would provide a much more convicingly like-native or better than native image. The character movement ghosting issue could easily be hidden with motion blur so it's hardly much of an issue.
Perhaps for those who actually like motion blur....
I feel like a lot of the problems in FSR versus DLSS can be explained by 2 things. One is that the FSR sharpening slider seems too aggressive, creating crunch and fizzle where it shouldn't be. The other is that DLSS is doing ML image analysis on the whole 3D frame, while FSR only does frame accumulation and reconstruction. It doesn't have a heuristic model to lean on when reconstructing, leading to the in motion pixel breakup and object fragmentation. It simply doesn't "understand" what objects should look like, so it's making a best effort decision about the pixels based on the frame data it does have.
DLSS, with the ML model, has a heuristic understanding of the frame and objects at a shape level. It can identify and reconstruct whole shapes in an image. By contrast FSR is only capable of understanding pixel history and motion vectors. It does not "know" what shapes are and as such can sometimes cause fragmentation of shapes when it tries to match pixels together.
To my understanding, that was the case for DLSS 1.0 but it's not true for DLSS 2.x. There is no acutual ML done while running the game and it also solely relies on motion vectors color data and depth data just like FSR. The reason DLSS 2.x is better is partly because the algorithm is more mature but mostly because it can use more samples per frame than FSR to reconstruct the image thanks to the increased speed of the tensor cores.
@@TheTaxxor I don't think learning ever occurred on the device, both versions use a neural net for sure.
@@TheTaxxor as far as I know both versions of DLSS use ML models computed offline. They aren't gathering training data in real time, but they are applying the heuristics of the machine learning model to the real time data being collected. DLSS 2.0 introduced a more broadly applicable model than was present in 1.0.
By contrast FSR 2.0 doesn't have a neural net or machine learning model at all. It's working exclusively on data collected in real time and hand tuned heuristics, rather than machine learned heuristics.
@@TheTaxxor that said, yes the tensor cores are a significant advantage. They are almost exclusively why DLSS works, because they can accelerate the application of the neutral net. Theoretically AMD could use a neural net in FSR, but the code would probably have to run on the CPU or less efficiently on the GPU AND consume some of the raster bandwidth in the process.
I hope "mediocre but universally compatible" doesn't win out over "superior but requires specific hardware," as tends to happen. If FSR and DLSS eventually become identical in performance and quality (similar to how everyone has good anisotropic filtering now), that would be cool though.
Completely agree. I wish consoles would use nvidia instead personally.
Fsr 2.0 is a big improvement but nowhere near as good as dlss. I think machine learning is just the way to go and focusing on things like fsr are the wrong direction personally.
Though I'd love to be wrong
14:56 fsr seems to break the sharpening of the grass behind the antenna thing
It doesn't brake sharpening but the reconstruction which create a pretty annoying pixelated effect that remind extreme sharpening, AMD traded ghosting with this artifact while the AI in DLSS is able to discard only the bad data and correctly reconstruct disoccluded pixels
So happy to hear this song again 😭
14:53
This is image ruining for me, look at the left side of the device in FSR. There's a weird trail of sharpening or something that looks really odd in motion.
I'm surprised it wasn't even mentioned in the video.
I was personally a bit blown away by how good FSR 2.0 looks in Deathloop recently on my 1080 Ti.
Good progress by AMD but dlss is still king of image reconstruction.
Well, they do have a bit of a head start
@@DanKaschel who cares if they have a head start lol. I want the best possible
@@NamTran-xc2ip what best? Are you playing games with 3x Zoom? The so called "imperfections" dont get noticed on blind tests.
It looks very promising, hoping we'll see a bunch more games use it shortly.