@@kraithaywire lol have you guys noticed that Google is having trouble with our spelling or slang and giving the option to 'translate to english' even though it is english. lol Both of you two have translate under your comment.
This reminds me of those old Xzibit memes. "We heard you like upscaling. So we put an upscaler for your upscaler so you can upscale while you upscale."
God those upscaling and framegeneration tech keep getting better, im looking forward to how it will be implemented in portable gaming, render a 720p 30fps game and present it at 1080p 60fps with minimal latency penaulty
NPUs are already major CPU features. Dedicated chips for esoteric functions are a very hard sell these days in the consumer sector. Even back when PhysX appeared it was never expected to be a big seller. My write-up of the CES that year where it was first publicly promoted was that this was plainly an acquisition target for a larger company that would subsume the functionality into their products.
GPUs already fit AI acceleration really well and recent generations include dedicated silicon for it (e.g. Tensor cores in RTX are NPU) Point of NPU as part of CPU is optimizing for power on laptops (not having to wake up dGPU, if laptop even has one to start with)
@niter43 So the things which Nvidia calls "tensor" cores are approximately the same thing as an NPU? Except that the NPU is part of the CPU silicon and the "tensor" cores are part of the GPU silicon? So could GPUs with enough of these cores add AutoSR support to a Windows system? Seems unlikely. I also think for major AI applications like LLMs, people don't even use Nvidia's "tensor" cores (or any NPUs) but rather their "CUDA" cores. Needless to say, AMD uses different names for all these things.
@@cube2fox there was just no (and probably still to this day) industry agreed term back when Nvidia introduced Tensor cores. Nvidia named them refering to TPU acronym, it was quite popular back then with Google calling their own new accelerator chips TPU. Just like industry didn't immediately came with "GPU" term. Nvidia's own LLM interference ("execution") engine uses Tensor cores. As seen in their "Chat with RTX" app or open source TensorRT-LLM library it's based on. Annoying to use, but seems to be the fastest engine for Nvidia hardware. llama.cpp/vLLM and others seem to use CUDA mostly for compatibility with older generations.
We already have 8k gaming, Nvidia added the Ultra Performance mode to their DLSS when they launched their Geforce 3090... specifically to market being able to run DOOM (2016) at "8K" resolution & acceptable FPS.
For what it is, it looks great. The more options there are, the better. I hope ms improves arm gaming performance either through software tweaks or the 2nd gen of these types of processors. The idea of the npu doing the work is an interesting one and i look forward to all the df coverage of it. Also, I've done no research but why is auto sr limited to 700-900p? Is the npu not up to the task or is there some other explanation?
Probably the NPU is not able to do more. The poor gaming performance for the Snapdragon Elite may be partly due to Microsoft's x86 to Arm emulation layer. However, it seems likely that the poor performance is mostly caused by Qualcomm's GPU driver, which doesn't contain two decades of game optimizations like Nvidia and AMD GPUs. Intel had similar issues with the driver for its then new GPUs.
10:10 AutoSR was way sharper here, which makes me think the stability issue was from Control's TAA. What if you had 720p DLAA + AutoSR vs 2160p DLSS Ultra Performance? That would've been a more fair test, so we're comparing the image with the same anti-aliasing
@@HybredNone of these technologies have a usable ultra performance mode, outside of the XeSS versions in intel GPUs, which turns to the same problem they have with DLSS. The point, I think, is that the comparison is unfair. It’s not to say “microsoft did a bad job here”, it’s to say “tame your expectations, this is not going to fix anti-aliasing problems as the other upscalers do”.
@@Dinjoralo. Why does it only work on a single processor? That wasn't made clear - I thought AutoSR comes on all CoPilot devices. It actually needs hardware to do that?
Like them or not, these are wonderful tools. I’ve always been disappointed they debuted & largely remained on the high end devices when it’s the underpowered (or low power) devices where they make the most sense. I’m hoping some of these really powerful APUs we are seeing will be able to leverage the NPUs AMD is adding to their chips for a ML-upscaler & framegen. No idea if the architecture allows it, but they are all just matrix multiplication accelerators (as I understand) so it’s the right math at least. My only peeve is that every company is remaking the same tech over & over, some standards & coordination would be good for consumers.
They make sense at every level. They actually work even better at higher resolutions with higher end hardware. I'm enjoying 4k gaming on a 4070 thanks to up-scaling tech. Sure it's not quite as good as 4k native but it's still better than 1440p native for the most part. There's still plenty of older and less demanding games I can run at native too.
That looks surprisingly good, and when you showed the 4-way performance comparison there, I couldn't believe it. This will be a godsend for all lower-powered mobile devices in our AI-powered future.
I launched Witcher 3 on my Surface Pro (11) and I was surprised about the clarity when I launched it. Was running at 50-60 FPS inside the city and the image looked really quite good on the 13 inch screen. The downside is that the UI elements looked quite blurry, but that's not that annoying really.
So, why can't this run on the GPU and be applied to browsers so we get upscaled videos (TH-cam, Twitch, etc)? I know Nvidia and AMD both have features to do that, but neither seem as good as AutoSR. Imagine upscaling 720p and 1080p videos to whatever resolution AutoSR upscales to and then using frame generation (like Lossless Scaling) to make them 120 fps.
Why? Probably no reason. The power of modern GPUs likely runs rings around these NPUs. How? Someone needs to be able to implement it, probably translating to CUDA for Nvidia or something special, and ROCm for AMD maybe?
If it is really post processing, a tech like this could be made into an hdmi dongle, to upscale anything, old consoles, switch etc., some hdmi dongle with an ARM processor.
Im interested in Auto SR for emulators. As DLSS/XESS is probably never going to be usable, for them due to all the vectors needed. FSR maybe is usable but Auto SR may be the obvious choicem
Also Auto SR will be far better than RSR or any other non ai scaler as it runs through a neural net unlike. Which will increase fidelity and fine details more so.
@@Gravy1255 Lossless Scaling works on anything, you can do upscaling with various algorithms including FSR, frame generation, or both together. All you have to do is configure the settings, select scale, then focus the window the emulator is running on. Make sure the window size is accurate to resolution the emulator is currently rendering at, and then it'll upscale that window to fullscreen borderless at your native output resolution. You can also set it to autoscale games with certain configs based on executable or window names
I admit I know nothing of this black magic, but I'd love to see it implemented into all Xbox consoles. Most Xbox One and Xbox 360 games run in the ranges of acceptable resolutions for this.
Since this is post-processing, does this mean it'll work on pre-rendered things? I'm thinking specifically about FMVs of poor quality and prerendered back grounds.
It probably works on everything whatsoever if it's really post processing. Similar to the upscaling a TV can do. But the machine learning model is very likely trained for upscaling 3D games, so I assume it won't work as well on, say, TH-cam videos.
Has anyone invented something like DLSDR for older games? The thing about older games is that they're not very demanding on modern PC hardware and their engines often aren't designed for high framerates because consoles aren't either. Pushing fidelity is the logical way to make use of my excess computer power in these games but I want to do it a little more efficiently than with pure render scaling.
DLDSR is handy for games that don't support SGSSAA. Otherwise SGSSAA produces slightly more stable results (or you can do 8x SGSSAA for super overkill internal resolution if your GPU can handle it).
Forgive me if I'm missing the connection, but dose Reshade offer upscaling options? I have never used it that much. I want to say it has some version of FSR.
Honestly... that looks really good. Like, for 720p upscaled to a clearer image, it's a very notable improvement. Appreciate the extra leg work of ensuring a clean image first to give the post process a fair shot.
System wide upscaling technology that you can switch on/off in any game you want sounds like a really good idea. Something like this implemented in handheld like Steamdeck would make it a killer feature indeed.
Steam Deck already has an implementation of FSR that can be used for any game and has had it for over 2 years. AMD has a similar feature in the Windows driver called Radeon Super Resolution, which is basically universal FSR (not to be confused with Virtual Super Resolution, which renders at higher resolutions and downscales to your output). Lossless Scaling also offers many different upscalers (including FSR) as well as universal frame generation for $7. I use Lossless Scaling almost daily to play games locked at 60fps at 120. It's not ideal for everything, but it works perfectly for something like Dark Souls, and I'm even using it for Once Human so I can play at higher settings (and since my framerate is a bit unstable anyhow even on lower settings with higher framerate, locked or not). The only thing I wouldn't use it in is a competitive FPS, but for most everything else, it's an amazing experience in my opinion
@@klementineQt On the Steam Deck FSR is used only as a 'scaling filter' with no other options than sharpness to tweak it with. Didn't even notice it since it was labelled like that until digging through once more. I also don't know if it's FSR 1, 2 or 3.1 implementation. It would seriously need some more work put into it like input/output resolution and the addition of AMD Fluid Motion Frames if possible. At this moment it's a mislabeled option buried under wrong category (it's upscaling tech, not a filter) and you cannot really tweak it, nor do you know even what implementation it runs on.
@@LurkingLarperIt's not labelled wrong because it is a filter. That's what FSR(1) is. There's no adjustable flags because your input res is always what your game outputs and the output is your native res. The only thing you can adjust is the strength of the filter which it does have.
@@zavier6362 I've never seen DLSS or FSR put under filters in any game or system before. It's always it's separate setting since it's upscaling. Filters are used in emulation or retro gaming scene. With upscaling you can always configure the input/output resolution which is also known as the quality of the upscaling ranging from ultra to quality etc. I don't see why these commonly used terms and settings need to be used in different ways under Steam Deck specifically. It just makes it messier and also the feature is lacking manual adjustment. Also what does this part even mean: "There's no adjustable flags because your input res is always what your game outputs and the output is your native res." There should be adjustable flags, that's the whole point of upscaling.
If they can make this work from higher res inputs & fix HDR support this could be awesome. I use Lossless scaling whenever 2160p DLSS isn't giving the performance I want, upscaling from 1800p or 1620p. But this seem leagues ahead of any of the algorithms on their so could be useful. It could also be a cheap way for Xbox to compete with PS5 Pro. I doubt they'll do this but imagine a Series X/S refresh with an added NPU. Games using FSR 2 could upscale to slightly reduced resolutions to limit artifacts then use AutoSR to go the rest of the way I.e. instead of 1080p to 2160p it could be 1080 to 1440p then AutoSR to 2160p, the image would probably be similar (though less artifact ridden like PSSR) whilst freeing up a bit of performance. And for the increasing amount of games missing 2160p and even 1440p this would be gamechanging. Again I know this won't happen but I think it hints at the future of upscaling, a mixture of heavier temporal techniques mixed with this new Ai based spatial method. This will be especially crucial if 8k does take off in the next generation as a technique like this is the only way a decent 8k image is happening (and at 8k the difference between this & a native render would be invisible on any screen less than like 120").
@@RedPillAlways💀 but true, they were trying to dissolve it in the xbox one days but just decided to burn it this gen with gasoline When I heard the marketing leader critise microsoft due to no funds for advertising and shareholders getting spooked I knew it was over. It had been getting slowly bleed by microsoft executives into a service.
I would like to see an RSR comparison: both are post processing filters that can be applied to any game. Once those new Ryzens are out, owners of those devices will probably have a choice between either AutoSR via Windows or RSR on a driver level, fitting a similar use case.
@@Nat-yf6ffWhy would that be a problem? Both are solution for the same problem, the fact that they have different approach only make it more interesting to compare really
@@danarseptiyanto4066 Not really. FSR is still implemented on a per-game basis and has access to more data than Auto SR does. Rich even explained in the video why it's not a fair comparison, but I suppose it's asking too much for people to actually watch the video before rushing to the comments section to bless everyone with their awesome hot take.
@@CaptainKenway There is Radeon™ Super Resolution (RSR) which is universal driver implementation of FSR 1, that is also strictly post-process with no input from game's rendering pipeline. As for comparriosion i dunno how well Lossless Scaling program works on ARM CPU but it also has implemented FSR 1 and NVIDIA's NID and has it's own post-process upscaling method
I think it would be interesting (assuming you can get them to run) to try out Auto SR on games that are too old for the other upscalers. Maybe some Unreal Engine 3 games? Some stuff from 2010-2015 ish maybe?
This "locked" 720p internal resolution makes me quite suspect of this being utilised for Microsoft's rumoured 4k handheld. If you had a powerful enough NPU I imagine you could reduce that 12ms latency somewhat, allowing them to push a 720p console with the same featureset as the home console, ray tracing and nice post-processing, and upscaling back to 4k.
@@ZeitGeist_TV I have a gaming laptop with gamepass..and cod I haven’t played since black ops 2. It’s better to purchase off steam since you are getting the complete pc version, and not a console port.
Technology has also invented "a pair of glasses" as a preprocessing filter for your eyes to see a sharper and more detailed. image. Auto SR feels like additional glasses
This looks interesting, useful for devices/games that need it. But some of the failures are surprising, for example the horrible noise around Jesse's shoulders at 8:52, or the complete failure to reconstruct thin power lines in CP2077 e.g. at 10:32. I wouldn't be surprised if Radeon's in-driver upscaler (RSR, based on FSR 1.0) wouldn't be very close, given that ML upscaling is apparently very limited without temporal data and other inputs.
I like it. I'd take this over FSR2's temporal artifacts I think. I'll stick with DLSS on my 48" 4K OLED but the AutoSR clips look fine on the 16" laptop screen I watched this on.
DirectSR also opens the door for future, better upscalers to work with existing games. It's been a long time coming. Thanks Microsoft for being able to force the industry to play nice.
I’m so curious with stuff like this how much the NPU or whatever it’s called is actually needed or how much more efficient it is than running it just on a CPU. Like, could they reasonably support it on non Copilot+ machines / machines without a dedicated neural processor or nah?
I don’t believe so, neural processing units are pretty specialized to perform low precision arithmetic and feature novel architectures. I assume it’d be like using a CPU to render graphics
Even with a relatively beefy NPU (Microsoft requires at least 40 TOPS) Auto SR only supports upscaling from 900p. Which suggests it isn't a particularly light task. So I guess something like an NPU is really necessary.
I wonder if this could be used for Video like RTX super resolution. It is a post process effect and seeing as many players already use DX11 it may be possible to implement. I don't see why an upscaler like this couldn't work & if it did it could be a gamechanger. Whilst it lacks the compression artefact reduction of RTX SR it certainly seems to be leagues ahead in terms of sharpness & clarity and it's real-time unlike Topaz, so for high bitrate 1080p content something like this could be incredible.
I believe Lossless Scaling doesn't have these strict NPU requirements as Copilot+. I guess Lossless Scaling doesn't require an NPU, nor that it even is AI based.
Lossless Scaling tries to scale low res images without the usual blur we get with other upscaling techs, it doesn't add new detail or tries to make the images look higher res. AutoSR does try to get an higher res image and adds new details (or fake pixels, if you prefer the term).
More of a traditional form of upscaling is how I see it. Like what you use for video playback perhaps. It's not a huge difference but it's a way to upscale lower resolutions to higher resolutions without it having upscaling problems because the render and output resolutions aren't an even multiplier like 2X. So you can upscale 1440p/1080p to 4k without is being a blurry mess easier. At least that is my understanding. I guess they just need to fix the artifacts.
If Auto SR produces its own anti-aliasing, wouldn't it be beneficial to turn off in-game anti-aliasing solutions in these games? Otherwise the algorithm has to fix the blur caused by TAA solutions. Also, if its possible to create a neural network that can do this, what's stopping anyone from making a neural network that can manifest its own guessed motion vectors?
AutoSR sounds interesting but DirectSR sounds like the solution we really need. Looking forward to when that releases and the big 3 make their SR solutions work with it.
I really like the approach of cleaning up the signal vs forced boosting to a fixed resolution. I think the quality of the image is very good especially considering the fact it has to work with so much less information vs DLSS. DLSS and FSR sometimes has the bad habit of adding shimmer and overly sharp artificial looking image, Auto SR seems to do the exact oppositie, it cleans it up and smooths it.
We seem to be getting to an interesting spot. There will come a time when resolution will be a meaningless property of performance with this level of "works everywhere" upscaling.
My big hope for NPUs showing up in commodity hardware was exactly this - low-compromise, vendor-agnostic AI upscaling. I can only hope that what seems to be a generic NPU-accelerated upscaler gets integrated directly into games and given the temporal treatment. Of course, that all relies on this shaking out to be vendor-agnostic once x86 NPUs have rolling out.
I'm wondering how this would look on old games. If it can take games that were made with a lower poly count and clean them up, it might be really nice.
Exciting. The benefit of having this be at the MS level and widely accessible, is that game creators can start to program for the strengths of this platform. Just as studios have started to consider how to implement DLSS, standardizing around a core upscaler maybe gets them to better TAA solutions and such that maximize the benefit of this. Perhaps.
This autoSR stuff would be perfect for the docked experience of these handheld PCs! Playing native 720p on the go and kicking in autoSR to upscale the image once it's plugged on the wall and docked outputting to a 4K TV.
Awesome tech this is. Even tho Switch itself cant do this I use 2 mClassic, gaming cable, cinema cable & the 4k gamerpro all daisy chained & I get the results of better upscale picture in 4k so I understand what this Auto SR is doing. Whenever Switch 2 comes out I'll try the same setup to see the difference.
Yeah, this will be amazing for handheld gaming. Funny enough, this is basically what FSR1 was. But so much better. I hope this level of improvement will not be stuck on windows platforms, android phones or a new Steam Deck would profit so much off this.
All this effort for MS to be able to push lower end ARM laptops down people throats, only for financial gain. I'm amazed that people are just accepting this... But hey - AI!! WOW!
Really quite exciting. The next Xbox will surely have some sort of NPU. could be another drive in back compat work. Furthermore, not auto SR per say, but a dedicated SR built into engines so developers can have more expensive assets and technologies whilst running at lower than 4k resolutions but absolutely nailing higher res becuase of the neural network
They need to implement something like this in Switch 2 so that it will look decent in docked mode when hooked up to a native 4k TV (unlike the current switch).
They have - it's called a Sony TV 🙂Being Japanese, Nintendo probably assumes that everyone has a Sony TV and the Switch looks amazing on it. But it is scary to see the Switch on a flagship LG or Samsung compared to a Sony.
They need to provide a standard way to access motion vectors in DX11/12 that is standardised and any upscaling can use it, including Auto SR. That way proprietary solution vendors would still be happy. Nevermind, that seems to be DirectSR you mention at the end.
This is precisely what I've been saying apple should do with M4 and Metal FX, so far they haven't used the CU units on apple silicon at all for hardware based upscaling. M4 is supposed to add way more for apple intelligence, hopefully they also add it to metalfx
Microsoft Copilot+ (which I guess includes Auto SR) is supported on all CPUs (SoCs, "APUs" etc) which have an NPU with at least 40 TOPS. AMD and Intel already announced their chips with NPUs similar to the one in Qualcomm's chip.
Never forget that in Digital Foundry's original video for Gears of War Ultimate Edition, Jon put on some weird accent that we never heard again
Had to go check…what was that!
I did not know this😂
wait... what? I gotta check that later 🤣
Liquid Jon
LOL! We need this as a viewer question on df direct.
“Ring the bell for allegedly instant notifications” - killed me lol
3:25 spitting bars dayum
LOL yeah hahahahaha
🔥🔥🔥
God damn Rich and his pudding
@@kraithaywire lol have you guys noticed that Google is having trouble with our spelling or slang and giving the option to 'translate to english' even though it is english. lol Both of you two have translate under your comment.
Drake has been real quiet since this dropped
This reminds me of those old Xzibit memes. "We heard you like upscaling. So we put an upscaler for your upscaler so you can upscale while you upscale."
God those upscaling and framegeneration tech keep getting better, im looking forward to how it will be implemented in portable gaming, render a 720p 30fps game and present it at 1080p 60fps with minimal latency penaulty
Inject all of it directly into SteamOS please!
@@Belfoxy If Lossless Scaling get a Linux port we wouldn't need it
I'll take that u. You don't need it .
This almost seems like we're soon back to dedicated cards for certain processing. In the past PhysX, today AI.
Well I'd imagine they stick with it on the CPU or add it to the GPU
NPUs are already major CPU features. Dedicated chips for esoteric functions are a very hard sell these days in the consumer sector. Even back when PhysX appeared it was never expected to be a big seller. My write-up of the CES that year where it was first publicly promoted was that this was plainly an acquisition target for a larger company that would subsume the functionality into their products.
GPUs already fit AI acceleration really well and recent generations include dedicated silicon for it (e.g. Tensor cores in RTX are NPU)
Point of NPU as part of CPU is optimizing for power on laptops (not having to wake up dGPU, if laptop even has one to start with)
@niter43 So the things which Nvidia calls "tensor" cores are approximately the same thing as an NPU? Except that the NPU is part of the CPU silicon and the "tensor" cores are part of the GPU silicon? So could GPUs with enough of these cores add AutoSR support to a Windows system? Seems unlikely.
I also think for major AI applications like LLMs, people don't even use Nvidia's "tensor" cores (or any NPUs) but rather their "CUDA" cores.
Needless to say, AMD uses different names for all these things.
@@cube2fox there was just no (and probably still to this day) industry agreed term back when Nvidia introduced Tensor cores. Nvidia named them refering to TPU acronym, it was quite popular back then with Google calling their own new accelerator chips TPU. Just like industry didn't immediately came with "GPU" term.
Nvidia's own LLM interference ("execution") engine uses Tensor cores. As seen in their "Chat with RTX" app or open source TensorRT-LLM library it's based on. Annoying to use, but seems to be the fastest engine for Nvidia hardware. llama.cpp/vLLM and others seem to use CUDA mostly for compatibility with older generations.
"And the pudding in question is actual gaming". There aren't enough t-shirts in the world for Rich quotes.
It’s time for the Frankenstein test. Let’s test an image fed through every upscaler at once! 8k gaming? 😂
We already have 8k gaming, Nvidia added the Ultra Performance mode to their DLSS when they launched their Geforce 3090... specifically to market being able to run DOOM (2016) at "8K" resolution & acceptable FPS.
8k soup
This is such a good idea
Too much gaming
@@Nicholas_SteelYou can't say you are playing in 8k if that thing is on. Not even 4k.
Stop being delusional
man those hud elements remind me of AMD's Morphological Anti Aliasing from back in the day, that post processing stuff
This looks a lot better, morphological filtering just blurred everything, this actually upscales and filters 2d elements
For what it is, it looks great. The more options there are, the better. I hope ms improves arm gaming performance either through software tweaks or the 2nd gen of these types of processors. The idea of the npu doing the work is an interesting one and i look forward to all the df coverage of it. Also, I've done no research but why is auto sr limited to 700-900p? Is the npu not up to the task or is there some other explanation?
Probably the NPU is not able to do more.
The poor gaming performance for the Snapdragon Elite may be partly due to Microsoft's x86 to Arm emulation layer. However, it seems likely that the poor performance is mostly caused by Qualcomm's GPU driver, which doesn't contain two decades of game optimizations like Nvidia and AMD GPUs. Intel had similar issues with the driver for its then new GPUs.
A comparison between RSR and Auto SR would be great, Which both technology works in post process and claims that could run on all games.
Within the limited scope AutoSR is shooting to cover, I'd say it does the job pretty well.
I hope they implement something like this in their consoles or at least their next console!
Nah they’re spending lots of money on this technology that’s mostly just for gaming in order for it to skip their next console
@@Chunnibyoubaka with the recent business decisions made by microsoft's xbox gaming div, i wouldnt be surprised
10:10 AutoSR was way sharper here, which makes me think the stability issue was from Control's TAA. What if you had 720p DLAA + AutoSR vs 2160p DLSS Ultra Performance? That would've been a more fair test, so we're comparing the image with the same anti-aliasing
You can't get DLSS on machines with AutoSR. DLSS requires an Nvidia GPU, AutoSR is currently only on laptops with Snapdragon processors from Qualcomm.
@@Dinjoralo. Then using FSR3 Native vs Ultra Performance, XeSS or TAAU would've been a better comparison. Surprised they didn't do this.
it does, but also looks it has more halos and artifacts. I'm loving these AI upscalers competition tho. The more the better.
@@HybredNone of these technologies have a usable ultra performance mode, outside of the XeSS versions in intel GPUs, which turns to the same problem they have with DLSS.
The point, I think, is that the comparison is unfair. It’s not to say “microsoft did a bad job here”, it’s to say “tame your expectations, this is not going to fix anti-aliasing problems as the other upscalers do”.
@@Dinjoralo. Why does it only work on a single processor? That wasn't made clear - I thought AutoSR comes on all CoPilot devices. It actually needs hardware to do that?
Like them or not, these are wonderful tools.
I’ve always been disappointed they debuted & largely remained on the high end devices when it’s the underpowered (or low power) devices where they make the most sense.
I’m hoping some of these really powerful APUs we are seeing will be able to leverage the NPUs AMD is adding to their chips for a ML-upscaler & framegen.
No idea if the architecture allows it, but they are all just matrix multiplication accelerators (as I understand) so it’s the right math at least.
My only peeve is that every company is remaking the same tech over & over, some standards & coordination would be good for consumers.
They make sense at every level. They actually work even better at higher resolutions with higher end hardware. I'm enjoying 4k gaming on a 4070 thanks to up-scaling tech. Sure it's not quite as good as 4k native but it's still better than 1440p native for the most part. There's still plenty of older and less demanding games I can run at native too.
That looks surprisingly good, and when you showed the 4-way performance comparison there, I couldn't believe it.
This will be a godsend for all lower-powered mobile devices in our AI-powered future.
I launched Witcher 3 on my Surface Pro (11) and I was surprised about the clarity when I launched it. Was running at 50-60 FPS inside the city and the image looked really quite good on the 13 inch screen. The downside is that the UI elements looked quite blurry, but that's not that annoying really.
It’s impressive, but I do notice the footage is a bit choppier with the tool
So, why can't this run on the GPU and be applied to browsers so we get upscaled videos (TH-cam, Twitch, etc)? I know Nvidia and AMD both have features to do that, but neither seem as good as AutoSR. Imagine upscaling 720p and 1080p videos to whatever resolution AutoSR upscales to and then using frame generation (like Lossless Scaling) to make them 120 fps.
I'm also wondering if there's any reason it couldn't work on video. Or.. On streamed games..
Why? Probably no reason. The power of modern GPUs likely runs rings around these NPUs.
How? Someone needs to be able to implement it, probably translating to CUDA for Nvidia or something special, and ROCm for AMD maybe?
Microsoft Edge literally has this feature with supported hardware
@@simonedeiana2696 yeah, people just like hating on Edge so they never are aware of all the features It has
@@ZedDevStuffAnd I use Edge Since Xbox one X and now Series X and order some things of Amazon!!
If it is really post processing, a tech like this could be made into an hdmi dongle, to upscale anything, old consoles, switch etc., some hdmi dongle with an ARM processor.
Im interested in Auto SR for emulators. As DLSS/XESS is probably never going to be usable, for them due to all the vectors needed. FSR maybe is usable but Auto SR may be the obvious choicem
Lossless Scaling works on emulators
There's already RSR if you have an AMD card
@@SogonD.Zunatsu all I can find for lossless scaling in emulators is frame gen type stuff. Taking a 30 game to 60
Also Auto SR will be far better than RSR or any other non ai scaler as it runs through a neural net unlike. Which will increase fidelity and fine details more so.
@@Gravy1255 Lossless Scaling works on anything, you can do upscaling with various algorithms including FSR, frame generation, or both together. All you have to do is configure the settings, select scale, then focus the window the emulator is running on. Make sure the window size is accurate to resolution the emulator is currently rendering at, and then it'll upscale that window to fullscreen borderless at your native output resolution.
You can also set it to autoscale games with certain configs based on executable or window names
If this type of Auto SR is in use for the Switch 2 for Switch 1 games, that would be amazing!
I admit I know nothing of this black magic, but I'd love to see it implemented into all Xbox consoles. Most Xbox One and Xbox 360 games run in the ranges of acceptable resolutions for this.
No neural processing unit in any current consoles, so I don’t think you can
That’ll likely be next gen
That'll probably come from mid-gen refresh/next gen
Or imagine an AI upscaler for old consoles, with component and hdmi input with no lag🤔
No but of Microsoft could allow fsr if AMD would implement it into an update for xboxs i mean its an x86 amd chip it could be done
You should try this in VR enviroment where that extra detail is going to make a huge difference.
Holy crap! Yes!
*Adding Microsoft Auto SR to DLSS on Dragon's Dogma 2 should rendera bit better appearance while sparing some extra resources.* 👍
Whoa there! Let's put windows into dark mode before you show those bright system dialogues. Love the content!
Since this is post-processing, does this mean it'll work on pre-rendered things? I'm thinking specifically about FMVs of poor quality and prerendered back grounds.
It probably works on everything whatsoever if it's really post processing. Similar to the upscaling a TV can do. But the machine learning model is very likely trained for upscaling 3D games, so I assume it won't work as well on, say, TH-cam videos.
You can try the Lossless Scaling, it works on every window
For Control you can use the modded version released by a Dev that includes upgraded DLSS and higher quality effects.
The series x also has 97tops of int8 ml inference. Just enough to run Copilot and Auto SR! Interesting......🤔
It would probably come with an update when ps5 pro release. But it will be fully supported in all games when next gen release. This always happens.
Has anyone invented something like DLSDR for older games? The thing about older games is that they're not very demanding on modern PC hardware and their engines often aren't designed for high framerates because consoles aren't either. Pushing fidelity is the logical way to make use of my excess computer power in these games but I want to do it a little more efficiently than with pure render scaling.
DLDSR is handy for games that don't support SGSSAA. Otherwise SGSSAA produces slightly more stable results (or you can do 8x SGSSAA for super overkill internal resolution if your GPU can handle it).
It would be interesting to see comparison with the LS1 upscaler in Lossless Scaling
LS1 is also a machine learning upscaler
I guess one of the devs Microsoft found out about Reshade and thought "huh, that's a pretty good idea" ;-)
Forgive me if I'm missing the connection, but dose Reshade offer upscaling options? I have never used it that much. I want to say it has some version of FSR.
Microsoft should produce a mClassic like device with this upscale software on it.
Honestly... that looks really good. Like, for 720p upscaled to a clearer image, it's a very notable improvement. Appreciate the extra leg work of ensuring a clean image first to give the post process a fair shot.
System wide upscaling technology that you can switch on/off in any game you want sounds like a really good idea. Something like this implemented in handheld like Steamdeck would make it a killer feature indeed.
Apparently, it's currently limited to Qualcomm Snapdragon processors.
Steam Deck already has an implementation of FSR that can be used for any game and has had it for over 2 years. AMD has a similar feature in the Windows driver called Radeon Super Resolution, which is basically universal FSR (not to be confused with Virtual Super Resolution, which renders at higher resolutions and downscales to your output). Lossless Scaling also offers many different upscalers (including FSR) as well as universal frame generation for $7.
I use Lossless Scaling almost daily to play games locked at 60fps at 120. It's not ideal for everything, but it works perfectly for something like Dark Souls, and I'm even using it for Once Human so I can play at higher settings (and since my framerate is a bit unstable anyhow even on lower settings with higher framerate, locked or not). The only thing I wouldn't use it in is a competitive FPS, but for most everything else, it's an amazing experience in my opinion
@@klementineQt On the Steam Deck FSR is used only as a 'scaling filter' with no other options than sharpness to tweak it with. Didn't even notice it since it was labelled like that until digging through once more. I also don't know if it's FSR 1, 2 or 3.1 implementation. It would seriously need some more work put into it like input/output resolution and the addition of AMD Fluid Motion Frames if possible.
At this moment it's a mislabeled option buried under wrong category (it's upscaling tech, not a filter) and you cannot really tweak it, nor do you know even what implementation it runs on.
@@LurkingLarperIt's not labelled wrong because it is a filter. That's what FSR(1) is. There's no adjustable flags because your input res is always what your game outputs and the output is your native res. The only thing you can adjust is the strength of the filter which it does have.
@@zavier6362 I've never seen DLSS or FSR put under filters in any game or system before. It's always it's separate setting since it's upscaling. Filters are used in emulation or retro gaming scene.
With upscaling you can always configure the input/output resolution which is also known as the quality of the upscaling ranging from ultra to quality etc. I don't see why these commonly used terms and settings need to be used in different ways under Steam Deck specifically. It just makes it messier and also the feature is lacking manual adjustment.
Also what does this part even mean: "There's no adjustable flags because your input res is always what your game outputs and the output is your native res." There should be adjustable flags, that's the whole point of upscaling.
Your videos are a breath of fresh air, always enjoy watching
12:02 for a snapdrgn its goooood progres
auto SR looks pretty damn amazing
I'd love to see how it compares to Sony's Reality Creation on their TVs.
It looks good but still has that typical "blurry, wormy" effect of upscalers, they can do a better job with the budget they have.
Wtf is it supposed to be? All the comparisons look the same to me.
It looks like total shit lmao
if we could lower the base resolution for better performance and then upscale to base...
👌
If they can make this work from higher res inputs & fix HDR support this could be awesome. I use Lossless scaling whenever 2160p DLSS isn't giving the performance I want, upscaling from 1800p or 1620p. But this seem leagues ahead of any of the algorithms on their so could be useful.
It could also be a cheap way for Xbox to compete with PS5 Pro. I doubt they'll do this but imagine a Series X/S refresh with an added NPU. Games using FSR 2 could upscale to slightly reduced resolutions to limit artifacts then use AutoSR to go the rest of the way I.e. instead of 1080p to 2160p it could be 1080 to 1440p then AutoSR to 2160p, the image would probably be similar (though less artifact ridden like PSSR) whilst freeing up a bit of performance. And for the increasing amount of games missing 2160p and even 1440p this would be gamechanging.
Again I know this won't happen but I think it hints at the future of upscaling, a mixture of heavier temporal techniques mixed with this new Ai based spatial method. This will be especially crucial if 8k does take off in the next generation as a technique like this is the only way a decent 8k image is happening (and at 8k the difference between this & a native render would be invisible on any screen less than like 120").
"We've been following AI upscaling for Years, DLSS, XeSS, FSR"
hol' up
I don't get it
@@icerink239 FSR doesn't use any kind of AI
@@EVPointMaster ahhh yes, slipped through my mind
I hope that Auto SR will be released on xbox.
Microsoft don't even like Xbox
Xbone is dead brother
@@stratoshd9043yes, you’re right, the Xbox One line of hardware is no longer manufactured.
Xbox and ps5 don't have SoC with neural processing parts. Best they can get is FSR...
@@RedPillAlways💀 but true, they were trying to dissolve it in the xbox one days but just decided to burn it this gen with gasoline
When I heard the marketing leader critise microsoft due to no funds for advertising and shareholders getting spooked I knew it was over. It had been getting slowly bleed by microsoft executives into a service.
What about a FSR comparison?
It's not a machine learning technique
I would like to see an RSR comparison: both are post processing filters that can be applied to any game. Once those new Ryzens are out, owners of those devices will probably have a choice between either AutoSR via Windows or RSR on a driver level, fitting a similar use case.
@@Nat-yf6ffWhy would that be a problem? Both are solution for the same problem, the fact that they have different approach only make it more interesting to compare really
@@danarseptiyanto4066 Not really. FSR is still implemented on a per-game basis and has access to more data than Auto SR does. Rich even explained in the video why it's not a fair comparison, but I suppose it's asking too much for people to actually watch the video before rushing to the comments section to bless everyone with their awesome hot take.
@@CaptainKenway There is Radeon™ Super Resolution (RSR) which is universal driver implementation of FSR 1, that is also strictly post-process with no input from game's rendering pipeline. As for comparriosion i dunno how well Lossless Scaling program works on ARM CPU but it also has implemented FSR 1 and NVIDIA's NID and has it's own post-process upscaling method
I think it would be interesting (assuming you can get them to run) to try out Auto SR on games that are too old for the other upscalers. Maybe some Unreal Engine 3 games? Some stuff from 2010-2015 ish maybe?
This "locked" 720p internal resolution makes me quite suspect of this being utilised for Microsoft's rumoured 4k handheld. If you had a powerful enough NPU I imagine you could reduce that 12ms latency somewhat, allowing them to push a 720p console with the same featureset as the home console, ray tracing and nice post-processing, and upscaling back to 4k.
No one ever said their handheld would be or target 4K at all though.
No handheld can game well at 4K in the slightest
After having 3 vita’s, and then went to switch…and now steam deck, I have yet to see how a Microsoft handheld can compete at this point.
@@TheL1arL1ar It works just be a means to have Game Pass and thus C.O.D in a portable form factor.
@@ZeitGeist_TV I have a gaming laptop with gamepass..and cod I haven’t played since black ops 2. It’s better to purchase off steam since you are getting the complete pc version, and not a console port.
Rich your Microsoft Store banner @2:11 is legit authenticated hairworks, no glitches bruz ;D Fellow baldy here ☺
Technology has also invented "a pair of glasses" as a preprocessing filter for your eyes to see a sharper and more detailed. image. Auto SR feels like additional glasses
I had to watch this with Lossless Scaling as the video footage felt like viewing 25 - 30 FPS :)
This looks interesting, useful for devices/games that need it. But some of the failures are surprising, for example the horrible noise around Jesse's shoulders at 8:52, or the complete failure to reconstruct thin power lines in CP2077 e.g. at 10:32. I wouldn't be surprised if Radeon's in-driver upscaler (RSR, based on FSR 1.0) wouldn't be very close, given that ML upscaling is apparently very limited without temporal data and other inputs.
I like it. I'd take this over FSR2's temporal artifacts I think. I'll stick with DLSS on my 48" 4K OLED but the AutoSR clips look fine on the 16" laptop screen I watched this on.
DirectSR also opens the door for future, better upscalers to work with existing games. It's been a long time coming. Thanks Microsoft for being able to force the industry to play nice.
I’m so curious with stuff like this how much the NPU or whatever it’s called is actually needed or how much more efficient it is than running it just on a CPU. Like, could they reasonably support it on non Copilot+ machines / machines without a dedicated neural processor or nah?
I don’t believe so, neural processing units are pretty specialized to perform low precision arithmetic and feature novel architectures. I assume it’d be like using a CPU to render graphics
Even with a relatively beefy NPU (Microsoft requires at least 40 TOPS) Auto SR only supports upscaling from 900p. Which suggests it isn't a particularly light task. So I guess something like an NPU is really necessary.
I wonder if this could be used for Video like RTX super resolution. It is a post process effect and seeing as many players already use DX11 it may be possible to implement.
I don't see why an upscaler like this couldn't work & if it did it could be a gamechanger. Whilst it lacks the compression artefact reduction of RTX SR it certainly seems to be leagues ahead in terms of sharpness & clarity and it's real-time unlike Topaz, so for high bitrate 1080p content something like this could be incredible.
720p resolution in 2024? Is this really the future of gaming?
Mr. Lead better, you are a very Intelligent man. Much respect to you and thank you for all that you do. Great video 👏🏼.
this can be amazing for handhelds
Isn't this the same as Lossless Scaling program, but running on chip dedicated to neural networks?
Both run on chips, both use NN
I believe Lossless Scaling doesn't have these strict NPU requirements as Copilot+. I guess Lossless Scaling doesn't require an NPU, nor that it even is AI based.
Lossless Scaling tries to scale low res images without the usual blur we get with other upscaling techs, it doesn't add new detail or tries to make the images look higher res. AutoSR does try to get an higher res image and adds new details (or fake pixels, if you prefer the term).
@@ghost085 Unless you're using bilinear upscaling, you're always adding new details by upscaling.
I really hope we can get this working with retro gaming... that would be amazing.
More of a traditional form of upscaling is how I see it. Like what you use for video playback perhaps. It's not a huge difference but it's a way to upscale lower resolutions to higher resolutions without it having upscaling problems because the render and output resolutions aren't an even multiplier like 2X. So you can upscale 1440p/1080p to 4k without is being a blurry mess easier. At least that is my understanding. I guess they just need to fix the artifacts.
THis would be so much better if it could use 1080-1440p image... Let's hope that when it arrives to AMD they will allow it...
Would have been interesting to compare Auto SR to NIS (Nvidia Image Scaler), these are both post proces upscalers.
If Auto SR produces its own anti-aliasing, wouldn't it be beneficial to turn off in-game anti-aliasing solutions in these games? Otherwise the algorithm has to fix the blur caused by TAA solutions.
Also, if its possible to create a neural network that can do this, what's stopping anyone from making a neural network that can manifest its own guessed motion vectors?
That flickering in the footage is horrible. Definitely would not use this feature.
I wonder how XeSS/FSR 3.1 Native (720p) + AutoSR will look like vs DLSS 3.7 UP 4K
AutoSR sounds interesting but DirectSR sounds like the solution we really need. Looking forward to when that releases and the big 3 make their SR solutions work with it.
I really like the approach of cleaning up the signal vs forced boosting to a fixed resolution. I think the quality of the image is very good especially considering the fact it has to work with so much less information vs DLSS. DLSS and FSR sometimes has the bad habit of adding shimmer and overly sharp artificial looking image, Auto SR seems to do the exact oppositie, it cleans it up and smooths it.
We seem to be getting to an interesting spot. There will come a time when resolution will be a meaningless property of performance with this level of "works everywhere" upscaling.
The almost fixed 12ms per frame makes me curious if and by how much the latency can be improved by future hardware. Cool tech, overall!
My big hope for NPUs showing up in commodity hardware was exactly this - low-compromise, vendor-agnostic AI upscaling. I can only hope that what seems to be a generic NPU-accelerated upscaler gets integrated directly into games and given the temporal treatment. Of course, that all relies on this shaking out to be vendor-agnostic once x86 NPUs have rolling out.
I'm wondering how this would look on old games. If it can take games that were made with a lower poly count and clean them up, it might be really nice.
Bro, your videos are gold.
A future where Direct XR allows consoles and PC to ship with games that support whatever AI scaler the end user wants will be nice.
The modern equivalent of weird scaling filters for old SNES games that got rid of the pixels by connecting them with diagonal fills.
Oh nice, Windows got Proton GE's FSR!
this tech should be great for downloading 720p videos to save space
No need to 2 pass h265 videos. Just keep them at 720p and then upscale. Data hoarders would be happy.
Exciting. The benefit of having this be at the MS level and widely accessible, is that game creators can start to program for the strengths of this platform. Just as studios have started to consider how to implement DLSS, standardizing around a core upscaler maybe gets them to better TAA solutions and such that maximize the benefit of this. Perhaps.
looking at this in a comparison between it and FSR1/NIS, it's the clear winner.
This autoSR stuff would be perfect for the docked experience of these handheld PCs!
Playing native 720p on the go and kicking in autoSR to upscale the image once it's plugged on the wall and docked outputting to a 4K TV.
Every modern TV has options like Auto SR and AFMF, they're just called different.
@@eSKAone- but does it look that good?
@@andremalerba5281 nah. Nothing can beat a NPU. Not those shitty tvs.
What about good old games with no UI scaling like Icewind Dale 2 or Fallout 2 meant to be played at 600p or 768p?
Lossless Scaling or Magpie
Awesome tech this is. Even tho Switch itself cant do this I use 2 mClassic, gaming cable, cinema cable & the 4k gamerpro all daisy chained & I get the results of better upscale picture in 4k so I understand what this Auto SR is doing.
Whenever Switch 2 comes out I'll try the same setup to see the difference.
Why the hell can't we feed it a 1080p image to upscale it to 4k?!
It would be interesting compering this to "nis" (enabled in control panel) which also works on a similar way minus the ai
Yeah, this will be amazing for handheld gaming. Funny enough, this is basically what FSR1 was. But so much better.
I hope this level of improvement will not be stuck on windows platforms, android phones or a new Steam Deck would profit so much off this.
All this effort for MS to be able to push lower end ARM laptops down people throats, only for financial gain.
I'm amazed that people are just accepting this... But hey - AI!! WOW!
Really quite exciting. The next Xbox will surely have some sort of NPU. could be another drive in back compat work. Furthermore, not auto SR per say, but a dedicated SR built into engines so developers can have more expensive assets and technologies whilst running at lower than 4k resolutions but absolutely nailing higher res becuase of the neural network
I think it is a very viable upscaling solution to play on smaller (up to 20 inches) screens. It looks way better than FSR.
They need to implement something like this in Switch 2 so that it will look decent in docked mode when hooked up to a native 4k TV (unlike the current switch).
switch will have the nvidia technology it will be far better.
Like jaimeguzman9086 said, DLSS is coming to switch 2 which will be the best tech available.
They have - it's called a Sony TV 🙂Being Japanese, Nintendo probably assumes that everyone has a Sony TV and the Switch looks amazing on it. But it is scary to see the Switch on a flagship LG or Samsung compared to a Sony.
With Nintendo using Nvidia Soc's I have no doubt the next generation Switch will support DLSS. I'd imagine DLSS 3+Frame Generation.
It looks very similar to what my Galaxy S24 does to images that are zoomed in.
Id say for a weak Snapdragon CPU on a laptop its pretty impressive, id be happy enough with it. Good work.
New Richard videos are always appreciated.
They need to provide a standard way to access motion vectors in DX11/12 that is standardised and any upscaling can use it, including Auto SR. That way proprietary solution vendors would still be happy. Nevermind, that seems to be DirectSR you mention at the end.
This is precisely what I've been saying apple should do with M4 and Metal FX, so far they haven't used the CU units on apple silicon at all for hardware based upscaling. M4 is supposed to add way more for apple intelligence, hopefully they also add it to metalfx
Maybe one day we'll finally get a great upscaler on consoles, so we can ditch 30 fps games or even make 90+ fps the standard.
720p DLSSp is THE hotspot for DLSS or any reconstruction. DLSSp, in original nvidia marketing slides literally said "designed for 1440p gamers."
Can auto-SR be run on tensor cores or only an mpu? Though I guess if you have access to superior DLSS this would be kind of pointless.
So can you use Auto SR on normal x86 desktops or only Qualcomms laptops?
Only Qualcomm because npus
You can probably use it on any pc with an NPU, but currently only Qualcomm devices have one.
@@SterkeYerke5555the horrible named New AMD laptops have then I believe.
@@mystic3309 strange that you think only qualcomm has npus
Microsoft Copilot+ (which I guess includes Auto SR) is supported on all CPUs (SoCs, "APUs" etc) which have an NPU with at least 40 TOPS. AMD and Intel already announced their chips with NPUs similar to the one in Qualcomm's chip.
I think this is very interesting! I really look forward to reading all the Anti-Upscalers cry.
The proof of the pudding / is in the tasting
and the pudding in question / is actual gaming
Would be great if VR could do this
I would absolutely not mind using it, when having limiting hardware. Looks good enough for sure!
Soon, all upscale technologies will feature configurable pass counts, ranging from 1, 2, and beyond, up to 69 and even 420 passes.
Will it work on an m.2 NPU or does it need to be done on the main SoC to keep latency down?