AMD FSR3 Hands-On: Promising Image Quality, But There Are Problems - DF First Look

so recently AMD dropped FSR 3 frame generation with two titles supported at launch for spoken and Immortals of avum now myself John and Alex saw demos at Gamescom though didn't actually play them so going Hands-On has actually been a bit of a treat first of all to verify that outside of controlled conditions AMD has done a good job here but secondly to get a handle on other stuff that we didn't get answers to at the time like what's really going on with vsync support uh latency mitigation vrr that kind of thing let's quickly review what frame generation is all about Nvidia kicked it all off with dss3 and in many respects FSR 3 follows the exact same principle um the next frame is rendered and the one beyond that too then with a combination of optical flow analysis informed via inputs from the game engine such as motion vectors for example Le an intermediate frame is then generated that slots in between the two standard rendered images frame rate then typically receives an extraordinary boost in my tests with FSR 3 in Immortals of avium on an RX 7900 xdx at 4K resolutions about 71% uplift in frame rate versus standard rendering I'm careful with the words used to describe the frame rate increase because similar to dss3 I don't think you can call it extra performance as such even though both AMD and Nvidia are likely to be using that term the game itself is still performing as it was without frame generation and in fact the extra calculations required to generate the intermediate frame have a cost of their own so one might even argue that frame generation reduces performance strictly speaking the output though what makes it to the player is visibly much smoother gaming more frames it looks like extra performance but it may not feel like it because well buffering up that extra frame incurs latency and adds to the response time as you play it can feel laggier Nvidia uses its reflex technology to claw back as much latency as possible while AMD has its own anti-ag and anti-ag plus Technologies ideally you'd be looking for the response time with frame gen to be the same as it is without so in our dss3 testing um to show you how it actually worked we had to resort to a bizarre OBS workflow to acquire 4K 120 FPS captures and really it's essential because High frame rate is where these frame generation techniques should be tested and how they should be used that capture technique was torturous in getting clean video to show you but technology has now moved on thankfully and 4K capture in this video is achieved with this the Yuan SC 750n full HDMI 2.1 support 8K at 60 FPS capture and yeah obviously 4K at 120 frames per second too I've been sharing various media using this card with our backers on the DF supporter program for months now but this is the first time we've used it in anger in a project so to speak first up let's go over the basics we've talked about how FSR 3 is very similar indeed to dss3 in principle but there are some differences difference number one FSR 3 is cross Fender AMD has recommended specs for its use but it's really just a compute Shader so if the game runs on your PC FSR 3 should run as well the issue is that anti-ag and anti-ag plus latency mitigations are on the driver level so that AMD only but you should be able to use Nvidia reflex if the game supports it if you have a GeForce card still thanks to AMD older Nvidia cars now have a frame generation option which is pretty great isn't it uh the next difference is that dss3 can run frame gen from any input so native resolution dlss xss even FSR 2 frames um D lss3 will generate interpolated frames from any kind of Base imagery you give it amd's FSR 3 is not so flexible it only works with FSR 2 upscaling or FSR 2 anti-aliasing finally dss3 doesn't officially support vsync but does work with vrr displays and also supports vsync off with pretty decent frame pacing AMD strongly recommends vsync on here and we'll get on to why shortly um but well I'm pretty sure that vsync off isn't working properly and I'm not sure that vrr works either let's talk about frame rate uplifts then this is just scratching the surface really because frame generators should be tested in both CPU and GPU limited scenarios across a range of kit hit but look this is just a first look video not any kind of definitive analysis still core I9 here 13900 K paired with my Rx 7900 XTX my Benchmark sequence is the beginning of chapter 3 in Immortals of avium I'm using FSR 3 at Native AA settings with the game on Ultra and yeah I'm using vsync for this for reasons that will become clear anyway frame generation is delivering an uplift of 71% in frame rate terms and moving on to a repeat of that test on RTX 1490 a more capable card we're still getting a 67% Improvement so why do benchmarks with resync active I mean that's not the proper way of doing things is it well I'll tackle that shortly but for now I want to go back to show you something I noticed in the frame time visualization on the XTX bench I decided to use native AA mode which is 4K resolution with FSR temporal super resolution used only for anti-aliasing purposes drastically improving image quality a heavy workload like this ensures we're under the 120 FPS threshold so the cap doesn't impacts the performance differentials but what's concerning here are the frame times it's clearer if you take out the non-frame Gen side of the equation and just focus on the FSR 3 capture the Cadence just doesn't look right and we shouldn't pingpong between such drastically different frame times now take a look at another sequence that isn't running frame generation but has a similar reported frame rate the switch to the 4090 is purely to match the performance level and simply to illustrate that this is how vsync is supposed to present the frame times on the bottom right there conform to what I would expect from that performance level with vsync in a 120 HZ container my takeaway here is the FSR frame gen seems to have frame pacing issues with vsync active that get worse the further you are away from the refresh rate of your display so really you should be tailoring your settings to get as close to your display refresh rate as possible and that could be tricky so I think this is perhaps why AMD says that a 60fps minimum base frame rate is a good thing for FSR 3 frame generation now this is a secondary Immortals of avium Benchmark I carried out the higher the non-frame Gen frame rate on the left there the fewer the frame pacing problems via frame generation but this sequence also highlights just how volatile performance can be from scene to scene and with that in mind I can't help but think that FSR 3 frame generation frame pacing needs a bit of work here it needs to be more consistent across the board but what about vsync off then well I don't think it's actually working properly here's a look at some capture again sourced at 4K 120 FPS you'll notice a persistent thin bar running up and down the capture now What's Happening Here is that frames that persist for a very short amount of time are surrounded by frames that persist for far longer those little frames there are about 150 pixels high or thereabouts uh within a 216p image I suspect these are our generated frames we're talking about a frame there that's persisting for less than a millisecond about 0.6 milliseconds the fact that you can see it is because what you might call a run frame is sitting within a 120 FPS video container there is a dramatic Improvement to the amount of frames being rendered but when so many of them are only being displayed for around 0.6 milliseconds they're actually becoming artifacts they're becoming problems here's how our tools visualize that green bars are new frame information black bars are frames that are identical to the previous one red bars are the runt frames and I think it's the generated frames that are being presented there so the placement of the generated frames looks good pretty consistent but their persistence basically makes them an irrelevance again I'm running the capture here at slow motion but even at full speed on a 120 HZ screen the Run frames are often very obvious in this particular vsync off Benchmark visualization the frame rate counter is still counting the frames but the frame time meter there filtering out almost all of those run frames bottom line then the FPS counter says you're getting an extraordinary booster frame rate however frame time on non-run frames shows that the four frames you actually see are fewer in number than the standard non-frame gen rendering I think this is the reason why AMD strongly recommends VC variable refresh rate vrr couldn't get it to work on my LG C10 there's a signal tagged as vrr but no variable refresh rate is being ped picked up at all my takeaway is that right now at least FSR 3 is vsync only the weird thing is that amd's docs do talk about vsync off support so I shared my findings with team red to get some clarity and here's their reply for the initial launch of FSR 3 we focused on laying a strong foundation for the development community and launching with the best user experience which meant having vsync on working well FSR 3 frame generation in our two launch titles were tested with both freesync on and off in combination with vsync enabled for vsync off use cases the FSR 3 implementation in our two launch titles for spoken and Immortals of avum is using a slightly earlier version of the FSR 3 codebase we have already identified improvements which have been rolled out to the GPU open FSR 3 code release which ensure better pacing between generated and realf frame presentations when playing the currently released titles on a high refresh monitor we recommend playing on high refresh monitors with vsync on okay so back to me now um the bit about testing with both freesync on and off in combination with vsync enabled just doesn't compute for me it suggests that vrr is working but it's definitely not for me I'm getting a vrr signal going through to the display but there's no actual variable refresh rate it's locked at 118 Hertz okay so I think we need to accept that frame gen is still in its earliest iterations we at DF had issues with dss3 at launch where vrr support was problematic if frame rate moved above the vrr window you'd get tearing and that's not the experience we expect from variable refresh rate that was fixed in due course thankfully these days you just select vsync on globally on the control panel and well hopefully the issues I've highlighted today with FSR 3 will also get sorted uh the Run frame problem with vsync off it's not an Immortals of Aven Problem by the way as here's the same thing happening in spor foken the issue being that the relatively even distribution of generated frames in avium is all over the place here I mean AMD is saying this is early code so let's focus on vsync on where across a lengthy segment of gameplay for spoken m maxed out in FSR 3 quality mode averaged 116 frames per second but yeah obviously the longer your clip and the longer you are at the frame rate cap the more the average is smoothed out still the fact is that a lot of the game does play out fully maxed at 120 FPS offset by the effects heavy combat which delivers a fair bit of frame time variance and can feel a bit choppy even so you're still getting big frame rate increases you just have to accept the vsync judder my takeaway is that you kind of need to Target your set refresh rate and adjust settings accordingly so you don't deviate too much from that figure I mentioned that earlier but it is tricky frame generation generally does a pretty good job of amplifying frame rate into the HFR High refresh rate window but vrr Smooths out the experience and for me at least that's the crucial component that just isn't working for me at least right now okay so I've not talked much actually I've not talked at all about the quality of the interpolator frames that's core to what Alex is going to be doing in his much bigger video but again FSR 3 Works in a similar manner to dlss 3 in fast moving scenes especially interpolated frames won't look as good as natively generated frames right but the thing is that they're always surrounded by two traditionally rendered images so artifacts if you can see them are kind of strobed therefore making them more difficult to notice lock a game to 120 FPS and each frame persists for 8.3 milliseconds which isn't a great deal of time anyway for the interpolated frame to persist now if you have a concertina effect of a lower quality frame surrounded by higher quality frames as I said it's harder to notice and that's one of the key reasons why frame generation is best suited to high refresh rate gaming where things come unstuck is if artifacts persist over a number of interpolated frames I'm going to leave the Deep dive analysis and stress testing to Alex in his content but what you've been seeing throughout this video is dlss frame gen based on dlss 2 quality frames compared against FSR 3 frame gen based on FSR 2 quality frames the sequences of Clips here are all locked to 4K 120 so perfectly alternating between native and generated frames I'm using the 4090 for dlss 3 imagery and 7900 XTX for FSR 3 another reason I'm not going to talk too much in comparison terms about the frame gen techniques on display here is that they're based on different quality inputs which makes like for like testing a lot trickier we'll need to work on this and try to get FSR 2 frame inputs working with dlss 3 to get a proper like for like matchup with the AMD frame gen Tech what I will say is that the most noticeable differences are coming from the upscaling methods right d lss2 versus FSR 2 as opposed to D lss3 versus FS s R3 both are on quality mode in terms of temporal super resolution but dlss has fewer artifacts so therefore fewer visual issues are going into frame generation I think my basic takeaway right now is FSR 3 is very comparable to dlss 3 and it's all done without custom Hardware on that front it's an extremely impressive achievement you can't really look at these side by side images in slow motion no less and instantly come up with a definitive winner right clearly FS sr3 in this scenario is getting the job done and it's the FSR 2 temporal super resolution component that is delivering the most obvious quality reductions against dlss on the left there I did some very quick latency testing in AVM and I used en video frame view to get the PC latency metrics which includes game logic and rendering but does not include peripheral or display lag Immortals of AVM supports Nvidia reflex meaning that markers a bakes in into the code allowing for latency to be measured even on AMD or Intel gpus which won't be running reflex of course some kind of bizarre results came up though with no lag mitigation in place at all measuring this static scene gave a 55.7 millisecond latency result without Fame generation rising to 64.5 milliseconds with Fame gen on I don't consider a Circa 10 millisecond lag deficit to be a big deal but this could change Accord according to content with basic anti-ag in play essentially four to 5 milliseconds are shaved off both frame gen on and frame gen off results so effectively we're kind of in the same ballpark but you know the reduction is there and it is quantifiable I'm sure you've skipped ahead to the anti-ag plus results and might be scratching your head and to be fair so was I in this scenario there's a massive reduction to latency without frame generation active I mean pretty impressive right a 28 millisecond reduction or thereabouts in latency compared to having no latency mitigation in place at all however weirdly with frame gen on anti-ag plus seemed to be consistently a touch less responsive than basic anti-ag I tried a couple of other situations in the game with similar results so it's a bit of a head scratcher really anti-ag plus potentially looks quite effective but in Immortals of avum at least it didn't really play ball with frame generation and so coming to the end of this video here's what I think of FSR 3 right now quality of the interpolation is fine rather impressive actually once you factor in that it's a software solution up against a hardware based solution from the competition and the frame rate uplifts are exactly where they should be on the hardware I've tested at least antiag plus I'd like to test on more titles but I was surprised to see it making little impact with frame generation in a Mortals of AVM where it could really make a big difference overall FSR 3 is a really interesting technology and I'm still quite taken aback by how this software solution is producing results that are up there against the competition but similar to the launch of dlss 3 if not more so actually a lot of things just aren't quite where they should be as a starting point though undeniably there is promise here so for a quick look I had a lot to say I guess and um we've not even touched on image quality really uh but I'm going to be leaving that to Alex as I said in the meantime I'm going to sign off so like subscribe share if you enjoyed the work and of course bell ringing allegedly gives you a notification whenever we quote unquote drop a new video so do consider that a DF supporter program Early Access bonus materials uh bonus videos DF direct weekly privileges an amazing engaged positive community and high quality video downloads of everything we do we're super proud of it and highly recommend you check it out but that's all for me for now on this one thanks for watching

Channel: Digital Foundry

Views: 202,350

Keywords:

Id: EBY55VXcKxI

Length: 19min 36sec (1176 seconds)

Published: Fri Oct 06 2023