Unreal Engine 5 DLSS 3.5 Ray Reconstruction: Nvidia's Vision of the Future of Gaming-

Nvidia Has released DLSS 3.5 Ray Reconstruction and Cyberpunk 2.0/Phantom Liberty has garnered all the headlines. But they also gave me access to an Unreal Engine 5 tech demo that shows off a ray traced Ramen Shop and offers an easy comparison of Ray Reconstruction on vs off vs native. We need to keep in mind that this is a demo produced by Nvidia specifically for marketing this feature, but even with that in mind it is interesting to play around with.
Test system specs (ResizeBAR/SAM ON):
CPU: Ryzen 7800X3D amzn.to/3Hkf7Qi
Cooler: Corsair H150i Elite amzn.to/3VaYqeZ
Mobo: ROG Strix X670E-a amzn.to/3F9DjEx
RAM: 32GB DDR5 6000 CL30: amzn.to/41XRtkM
SSD: Samsung 980 Pro amzn.to/3BfkKds
Case: Corsair iCUE 5000T RGB amzn.to/3OIaUsn
PSU: Thermaltake 1650W Toughpower GF3 amzn.to/3UaC8cc
Monitor: LG C1 48 inch OLED amzn.to/3nhgEMr
Keyboard: Logitech G915 TKL (tactile) amzn.to/3U7FzA9
Mouse: Logitech G305 amzn.to/3gDyfPh
What equipment do I use to make my videos?
Camera: Sony a6100 amzn.to/3wmDtR9
Camera Lens: Sigma 16mm f/1.4 amzn.to/36i0t9t
Camera Capture Card: Elgato CamLink 4K ‎amzn.to/3AEAPcH
PC Capture Card: amzn.to/3jwBjxF
Mic: My actual mic (AT 3035) is out of production but this is a similar mic (AT 2020) amzn.to/3jS6LEB
Portable Mic attached to camera: Rode Video Micro amzn.to/3yrT0R4
Audio Interface: Focusrite Scarlett 2i2 3rd Gen: amzn.to/3wjhlad
Greenscreen: Emart Collapsable amzn.to/3AGjQXx
Lights: Neewar Dimmable USB LED amzn.to/3yw4frD
RGB Strip Backlight on desk: amzn.to/2ZceAwC
Sponsor my channel monthly by clicking the "Join" button:
/ @danielowentech
Donate directly to the channel via PayPal:
www.paypal.com/donate?hosted_...
Disclaimer: I may earn money on qualifying purchases through affiliate links above.

[@danielowentech]

In the video I say this is path traced. However, I looked over the documentation from Nvidia afterwards and noticed it says: "this demo includes updates to ray traced effects like RTX Direct Illumination, Lumen Global Illumination, Lumen Reflections". It never specifically mentions path tracing. So it looks like that was a mistake. Apologies!

[@jannegrey593]

understandable

[@Tigerhearty]

Then i only need to wait for a generation of gpu that can finally do 60 FPS at 4K again with path tracing enabled while dlss 3 and 3.5 turned off before upgrading, thanks for the help.

[@dakelpoandgames2593]

Hey dude, the instability you noted with pathtracing and with dlss on happened in cyberpunk for me. If you have a 3070, I am asking if you can go to where Vicktor, the ripperdoc's shop is and turn on pathtracing and then also dlss. At 1080p or 1440p I was getting huge amounts of noise and instability on everything. Ghosting was prominent and the characters skin would have that noise you showed on reflective surfaces. Even the hairless cat had like a pink and purple shine to it. Just observe it yourself if you have the older update before 2.0. Its crazy how bad it was. I did end up upgrading to a 4070ti but the issues existed until this update.

[@dakelpoandgames2593]

Also, awesome video dude! Keep it up!

[@MegaLol2xd]

So RR works with normal RT, why not give it to people with rtx 30xx users who want to play cp2077.

[@guntersztajner7204]

I played Fort Solis, which is a UE5 game with Lumen, and there was a lot of shimmering with reflections and global illumination due to poor denoising. The solution was to increase the number of rays cast, but this had a very negative impact on performance. Here Ray Reconstruction not only fixes this, but even improves performance, crazy. I hope this amazing tech will be added to the dlss plugin for UE5 so that any future UE5 game can easily implement it.

[@ratharos]

Wait did they solvE that in fort solis? I touht it was a FINAL gather issue. I stopped playing because it was distracing

[@guntersztajner7204]

@@ratharos I fixed GI shimmering with this command "r.Lumen.ScreenProbeGather.TracingOctahedronResolution 16" ( default for highest settings was 8). It's around 20% peformance hit if I remember right. As far as i understood it increases the rays (samples) per pixel. Other way is to use "sg.GlobalIlluminationQuality 4" but the performance hit is way too high to justify it. I used Universal UE5 Unlocker to enable the console but you could probably just write these commands to engine.ini.

[@damara2268]

I think nvidia already said they are working on a RR plugin for UE5 so probably thats gonna make it easy to replace lumen denoiser with this one

[@ionseven]

Marketing bs. They don't want to sell us the stronger hardware at reasonable prices. DLSS was supposed to help with incapable hardware or prolong the life of older cards. Now we are right back to upgrading anyways, but instead of hardware we are getting GPUs for these software gimmicks.

[@lau6438]

@@ionseven Ray Reconstruction is available for all RTX cards, not just the 40-series.

[@leotide1990]

It pains me to say it with the prices Nvidia is charging these days, but... my goodness, ray reconstruction really is almost magical. This is amazing

[@NostalgicMem0ries]

thats why they price it so high, their stuff is superior and works, while amd just now to be introducing frame gen, nvidia had it for a year now, and not just that dlss also far superior to fsr, now with this ray reconstruction ray tracing is not even a battle anymore...

[@linhza501]

@@NostalgicMem0ries If they can lower the price with such good tech like this, they can gain even more grounds against AMD.

[@filip9587]

Nvidia is the new Apple. Selling an item known to be superior in Tech for absurd prices. Except Apple hasn't been the leader in Tech for years. Edit: Plus there's the brand name.

[@adriancioroianu1704]

​@@linhza501They don't compete with AMD. That chapter ended a long time ago.

[@robbieclark1784]

At $550, the 4070 is the best single player graphical fidelity focused card on the market except the 4090.

[@adriankoch964]

It's still mindblowing how we have near tangible real time raytracing now. Glasses not looking stable in the distance (and generally not interacting correctly with light) is imo very much preferable to "each frame taking an hour to render" as it should be considering to how ridiculously expensive it is to correctly render glass in computer graphics. (See any video for Blender about optimizing glass, as those either tend to turn black or take forever to render)

[@Crittek]

Imagine what DLSS 4 will be like. DLSS as a form of anti aliasing was completely broken 3 years ago.

Benchmarks and Games should include a photo mode where a single high res frame is rendered in natively (no upscaling) with crazy amount of rays, even if it takes a minute or two to render. It would give us a reference image and make it easier to compare to real time renderers.

[@KUZY2X]

That is also the reason why they do not do it. If potential customers are able to compare the native resolution with the upscaled they will definitely see that the upscaled version is not as good as Nvidia states. Personally, I'm very disappointed in this trend of using upscale instead of native resolution.😟

[@ZenXIV]

@@KUZY2X Do you understand how hard path tracing is to run? A 4090 can struggle to do path tracing in CP2077 without DLSS at any resolution. You would probably be hard pressed to even reach 60 at 1080p. GPU hardware is starting to hit a standstill, to get more power they are having to pump more power into the cards. Yes they might be able to squeeze more power without increasing power consumption but there will eventually come a time where they cannot make a card anymore powerful without cranking power limits higher. Moore's Law may be obsolete by 2036. Path Tracing is so insanely computationally intensive that it's a damn miracle that we can even do it in a real time. You want real time reflections & light behaving realistically then you have to put up with using upscaling tech like DLSS as it is simply unfeasible without it. Movies studios use farms of top end GPU's to render a single screen of animated films and even that can take fucking hours upon hours (Mind you they do this at high resolutions but still). DLSS is a fucking wonder of a tech, it can give you almost native quality of image while increasing your frames and unless you go looking for issues you will never really be able to tell the difference. It's working smarter rather than harder. Throwing tons of power at the problem is the only way you could ever do native resolution with path tracing, it's the brute force method. If you don't want path tracing or ray tracing in your game then cool stick with Rasterization that will inevitably hit a stand still of graphical quality, the argument of well Rasterization looks fine right now is a dumb one when you realise rasterization will never be able to achieve real lighting or proper reflections and don't say anything like well we have Lumen. Lumen is ray tracing, it's software RT. RT & PT is the future of gaming and it is pretty much impossible to achieve with native resolution because they are so demanding that hardware has to work smarter to keep up with it rather than brute forcing it's way to playable FPS. However currently developers are using the tech to forgo optimizing games but as with everything people do sometimes use it poorly but that doesn't make that tech bad. Nanite doesn't mean developers shouldn't optimize raw meshes and the way Unreal said that developers could just put in a mesh easy with nanite gave the entirely wrong perception, after all you don't need insane polycounts on simple items. Nanite also needs to be able to smartly reduce and increase polycount based on the user's FPS. You should be able to give it a low end and high end of polycount for a item so developers can control the limit of the items lowest poly count. These techs are still in the infant stages and lots can be done to improve on them, they will only get better with time. Dynamic LOD's and RT/PT are the future.

[@webx135]

2:20 Ohhh, that isn't smoke. It's steam. Steam from the steamed clams we're having. MMMM steamed clams!

[@danielpindell1267]

Frame times are better with it on as well which is impressive.

[@VeryFunnyStuffs]

Awesome video and comparison! On an other note: serene underlying sound! Has/adds something nice to it!

[@cheekarp2180]

Very good examples and explaining, thank you.

[@edvinjohnson4136]

If anyone's curious, the name of the background music is Mystic Maze from the album Puzzle Maestro: Casual Scenes & Ambience. Lovely little track.

[@offroadskater]

12 frames on a 4090 means we are at least two generations away from general population having access to the needed horsepower. Even the 1440p at 25 base doesn't change that timeline. 2 generations of nvidia for ray-tracing means 3 for amd. That adds up to 6 years for PC mid range cards. It will be tough for consoles in 2028 to have it. Maybe in dynamic resolution, but not in 4k. I'd be very ok with that.

[@CharlesVanNoland]

I believe that path-tracing will become more popular as they engineer faster algorithms for building and traversing scene hierarchies, maybe mixing it with other representations besides triangles, maybe using KD-trees instead of conventional bounding volume hierarchies. Who knows! Moore's Law is on the verge of death, with how much more expensive ASML's machines are becoming just to get down to smaller transistors - it's becoming prohibitively expensive. The ROI doesn't work out in favor of chip designers, unless they figure out how to make borophene or graphene semiconductors at scale, instead of using stupid silicon. We'll have to get smarter about how we are calculating our lighting if we're going to be stuck with crappy power-hungry silicon because there's not much more room for making the hardware faster without making GPUs consume 1000 watts.

[@adriankoch964]

"Photonic computing or bust"

[@ruseruser2227]

Not knowledgeable on the in-depth constraints there, but architecture optimization could become the next holy grail now that its becoming harder and harder to just get brute force improvement through smaller process manufacturing and higher transistor density. We're at that ugly middle point between the competition of traditional rasterization hardware and raytracing hardware on the same die, and ray tracing may slowly but surely phase out or switch places with rasterization to become the dominant hardware components on the die, resulting in further growth independent of Moore's Law. Really excited to see how the developments go though, it's astonishing seeing how far things have come in just the last 5 years!

[@racoonchief]

and how far into the future do you think it will be popular? what makes you think there won't be something better by then?

[@racoonchief]

It's a waste of time to ''believe'' in these sort of stuff.

[@ionseven]

Marketing bs. They don't want to sell us the stronger hardware at reasonable prices. DLSS was supposed to help with incapable hardware or prolong the life of older cards. Now we are right back to upgrading anyways, but instead of hardware we are getting GPUs for these software gimmicks.

[@aikanikuluksi4766]

This seems to demonstrate that: - Ray/path tracing is inherently unstable due to the small number of rays/paths and needs denoising. - DLSS (and likely any upsampling) amplifies the instability and increases the need for denoising. - Ray reconstruction does temporal denoising on RT+DLSS (effectively temporal averaging), resulting in a nice static scene. - Dynamically changing but local and unmoving picture elements may or may not have some artifacting. What the demo is lacking is major moving elements with notably different colors and lighting (such as walking NPCs) or a left-right moving viewpoint, which both seem to create notable artifacts in Cyberpunk due to the lag/smearing introduced by temporal denoising. As temporal denoising is required to properly even out the small number of traced rays, the question of interest (and not addressed by the demo) is whether it is even possible to eliminate such artifacts using consumer grade hardware.

[@Antagon666]

Also high frequency high contrast detail tends to produce interference patterns even when static.

[@Mat-hr1dg]

You are either very mistaken or missed some words, namely "in it's deployed form" from that first point. Ray tracing isn't inherently unstable and isn't limited by rays because of raytracing itself but rather because current hardware doesn't allow for faster calculations that would make more rays and bounces possible yet. What's so beautiful about raytracing though is that it perfectly scales with core counts since you can just average the areas which is basically what the denoiser is doing already. That means raytracing is a very scalable technology in its core and will get better quickly after a certain threshold of rays and bounces has been reached. Currently we are still so low on them, that changes in the demoise algorithm will show much quicker no matter if we can add a bounce more/get 2x the rays or not. When we achieve 4x-8x more we might get into a territory where the denoise algorithm will have to do a lot less guesswork and the static image will appear stable no matter what. Considering the tech, this will be sooner rather than later.

[@GKSchattenjaeger]

Is this tech demo publicly available for download? Can't find it.

[@guitaristkuro8898]

Is there a download for this somewhere? I can't find it. I don't want this to yet another demo Nvidia says you can download, but then they never release.

[@Verpal]

Honestly there are too much ghosting with RR now, however, just like DLSS NVIDIA have a history of constant update and improvement, it will probably become better in a few iteration.

[@Pwnag3Inc]

That you will have the pleasure of using once you buy a NEW graphics card. Lol. I dunno how long they can keep this grift going, but by god jensen is damn sure going to find out.

[@Tee_B]

​@Pwnag3Inc You mean like how Reflex is usable on 9 year old cards? How DLSS 3.5 is usable on Turing cards, despite it just having come out?

[@Tigerhearty]

RR is based on algorithms provided and trained by nvidia themselves so it can only improve over time with better and more refinded training of the model used by the ai tensors

[@Tigerhearty]

​@@Pwnag3Inc when we are all back to 1080P 60 for path tracing, in 5 year, it will be time for cosmic ray tracing, as light obviously depends on the refraction of atmosphere and light bouncing off of planet to enlighten the earth in indirect light, bringing you back to 720p 30 FPS with a RXTX 7090 ti

[@dafyddrogers9385]

​@@TigerheartyYou aren't forced to use it, the game looks good on raster and you can get 200fps

[@joshuas8912]

The artificing @ 11:40 may be related to how DLSS works in low light. DF talks about it a bit related to NPCs and how this was also the case in early versions of DLSS 2.0

[@Wobbothe3rd]

Not just DLSS but denoising in general. Quake 2 rtx does this woth no DLSS

[@ISAK.M]

Impressive, maybe I can use it in 5 years when I can afford a 4090

[@christophermullins7163]

"Flagship in 5 years" gang stand up!

[@McLeonVP]

in 5 years you would afford a RTX 6050 and will be the same xd

[@Seacat17]

Now I want to download this demo to play with it myself.

[@MoireFly]

The problem specifically with using techdemos for stuff like this is that it's hard to validate how well the baseline was chosen. DLSS looks better... but better than what, exactly here? Was the non-DLSS denoiser otherwise a reasonable cutting edge choice, or was it some poor fallback? Not trying to criticize DLSS RR here; it sounds excellent - but I think stuff like cyberpunk 2077 is a better demo here because we know what the alternative was; clearly the non-DLSS denoisers were not chosen carelessly or even specifically to make RR look good. That's less obvious here.

[@Uproar3323]

I would like to try this demo. Is there some place from where I can download this demo?

[@RazielAU]

Ray reconstruction is part of the same upscaler that DLSS 3.5 uses, it just includes that denoiser into the same AI model that handles the upscaling. In other words, it shouldn't negatively impact performance as it's all part of the same AI model that runs, it's just an alternate implementation that happens to be more efficient than the older DLSS path (DLSS on but ray reconstruction off = older DLSS code that happens to be slower).

[@Kogiz]

Random question but somewhat relate. Do you think that, when FSR 3 comes out, it will be usable with Nvidia Reflex + DLSS?

[@CyberStrike]

It is not possible to use the upsampling techs , DLSS/FSR/XESS, simultaneously during the rendering process.

[@L77kim77l]

Kind of hypnotizing: the mix of graphics, music and you talking...

[@marsovac]

You are not getting sparkles because the AI is trained to remove high contrast low surface area reflections. It usually means removing shimmering, but in many instances it also removes proper reflections, like those on the edges of floor tiles, and since it is AI it does some weird random artifacts at times (that was probably not provoked by the fan moving, but just by how AI works).

[@ishi...]

would have liked to see 1080p performace +RR +FrameGen vs native 1080p

[@pixel1145]

But what version of Unreal did they use? There is no RR in the DLSS plugin and the current one still doesn't support 5.3

[@chintoki]

Their own fork probably.

[@petyavodolaz]

Could you add a link where it can be downloaded? Google didnt really help :(

[@KhazAlp]

Can you provide the download link mate?

[@dmitriibob]

Is this demo available for public?

[@technologicalelite8076]

11:31 That and the stove in the back round as well as 2 spots torwards bottom left and bottom right of the screen. Weird!

[@Pamani_]

2:19 from the steam clams we're having

[@shayeladshayelad2416]

Lowers the ms and watt and giving better pic quality wow this is great!!!!

[@huaky24]

The type of Instability in the demo is what I didn't like in control, but control itself was a pretty good game

[@ickerolig]

Question is, does DLSS make nvidia cards more priceworth? I'm planning a build right now, and I can't really settle on whether to go for an amd or nvidia gpu.

[@Fantomas24ARM]

Yes. DLSS is significantly better than FSR in every way. At this point I think it is really a selling point.

[@bitbat9]

Absolutely. Frame gen on the 4000 series cards is amazing too

[@ladyinwight]

its one of the only real reasons to buy nvidia over amd right now, so yes it makes it more worth the price. but if you mostly play multiplayer games or dont particularly like upscaling then you'll get more out of a 7800xt right now. higher end nvidia cards also run insanely better than AMD when raytracing (i'd say 4070 is a good starting point for raytracing, lower than that and performance won't be great)

[@NadeemAhmed-nv2br]

It depends, at 4K there a tie but at 1440p and 1080p, dlss wins hands down, however if you have the RT power and your running at 1440 p.m. 1080p, then dlss looks worse than native

[@ickerolig]

So. As I'm someone that's content with a 24'' monitor and would never even think about going above 1080p DLSS isn't that big a deal? EDIT: Nvm that statement, the reason I'm building a new rig that can play more than Quake 3 and Tekken is because I want to play upcoming simracers, an ultrawide will probably tag along the new PC.

[@lokeung0807]

It make Nvidia’s card more attractive. Although it is not perfect, Nvidia proven they can improve DLSS well. It is very hard for Intel or even AMD to catch up.

[@mariop8101]

I don't see any DLSS, I only see one RT feature, for path tracing.

[@PandaP-]

Idk XESS in the short +-2yrs its been out, is better than current FSR. And is on par with DLSS 2. Not to mention, Intel is busy with drivers. If Intel were to make XESS2, either for arc or battlemage, it most likely be on par with DLSS3.1 as they have been working on a RR equivalent

[@steel5897]

Intel has a chance, they've been investing a ton in ML and RT from the start. AMD has no chance, they've been coasting on their console contracts and are now hopelessly stuck in the past when it comes to technology.

[@sherazdotnet]

@danielowentech how can one download this demo to try out?

[@rozzbourn3653]

are the weird artifacts being caused due to a reflection of something else on the shiny pots? edit: nm, you mentioned it later on.

[@sociofall]

yes, that's what I hoping for! does they actually inject ray reconstruction into Lumens pipeline to nihilate that's noisy artifacts?

[@Wobbothe3rd]

It uses Nvidia's own direct Illumination combined with Lumen global illumination. So it's not exactly full path tracing, but still requires denoising.

[@MasterGeraltofRivia]

The worse picture quality with DLSS On without RR isnt caused by DLSS itself, but by the fact the ray count is based on internal resolution.

[@manuelrodriguezcarmona3988]

I just got 6950xt, good buy? Im worried about this.

[@luckyyluck]

where can I download this?

[@zbuuu]

How to download this demo?

[@zaselimgamingvideos6881]

From where can i get the demo?

[@slay3rsaber]

Is this damo publicly available? I would like to check it on my system.

[@DoubleAAce]

Is there a way for us to see this demo on our pcs or has Nvidia sent it for you to show us?

[@danielowentech]

Nvidia sent it to me. I don't know if there is a public download. It's called Ramen Shop.

[@arenzricodexd4409]

Usually nvidia will releass their tech demo later.

[@OneTwoMark]

I dont understand why ray construction wasn't directly worked into DLSS? Whats the point of having yet another setting to enable, making it more confusing for a novice on what to turn on.

[@AR-ey1ur]

At what point did he check frame generation?

[@Nobody-sp7ug]

I mean it's a denoiser, cause thats what it does, but I dont think think you can really categorize it with denoisers like sigma, relax reblur and stuff. It's tensor cores running sparse inference on a trained ai model again, so it's the same as dlss image reconstruction in that way, but it's ai model is trained to recognize good ray trace results AND bad results, and pick out good results and toss bad results. It also moves denoising off the cuda cores, and puts it on the tensor cores, which while incurring a time cost, I mean, it HAS to have a time cost, the only thing thay doesn't is nothing.... but the time cost has more concurrency now between cuda cores, ray trace cores, and tensor cores. And this is just the first commercial generation of this ai model. Also, ray reconstruction was implemented in a full path raytracer in rtx overdrive first to show off, but it can be implemented in standard rtx ray tracers, and Nvidia said that's coming next.

[@WbosonLP]

Yes, great they put the AI cores to work now. 3 gens later, and then giving themselves a pat on the shoulder. I find it amusing. Anyways I´m more curious about why is the RR ON\OFF tied directly to DLSS? Can´t it be done separetley? And can´t it be ran on other HW than Nvidias? Also if the RR needs 5x the resources from Nvidia , isn´t it more tangible for them in the FUTURE to just abandon the project just to to use computing power on their servers to crunch our personal info for profit? I mean first it was all about RT will be super easy to implement- just script the ON\OFF toggle in. But now we have RR which requires 5x the computing power from devs\HW vendor and then the result isn´t perfect still... I´m still waiting to see what the future brings us.

[@Nobody-sp7ug]

@WbosonLP inventing graphics uses for ai and then successfully training the models to provide the desired result takes time. Also stop giving them supervillian ideas!!!!! I think you might be misinterpreting what nvidia said about 4x longer than dlss though. These are two different pipeline, one of which (cuda cores) takes a lot more render time than the other (tensor cores) DLSS /tensor cores werent doing anything raytracing related before, besides standard non tensor data type use, for data types the cuda cores don't have (cuda cores only have fp32 and int32) so the original time for raytracing involved stuff on the tensor cores was basically 0. The calculations and intersections were done on the raytrace cores, and the denoisers were run on the cuda cores, after dlss created a 4k image. They kinda fought each other. Ray reconstruction removes denoising from the cuda cores, and puts ai 'denoising' as part of the dlss pipeline on the tensor cores. Because thats basically what it's doing now, dlss for raytracing, except instead of ai reconstructing a normal looking low res image like dlss, it's ai reconstructing that weird 'ray trace view' which looks like it is just a bunch of dots. It's ai making new dots without any new rays needing to be calculated, and removing dots it thinks are bad results, based on its trained ai model, and the image dlss reconstructed from the low res render. Thats what makes dlss take 4 times longer now. But it's on the tensor cores, which are 8x, 16x, and 32x faster than cuda cores for this kind of work, and which can do most of their work side by side and independant of the cuda cores. Currently ray reconstruction adds no time, or gets time back over cuda core denoisers, or in few cases like rtx overdrive, barely takes away like a frame or two. Which...honestly... isn't really adding up. This is a win win for tensor and cuda cores, it should always get performance back, if it just involves those two. so that makes me think the difference thats taking up more time is in the ray trace cores. I wonder if they are taking on a larger load than before turning on rr in overdrive because the rest of the pipeline got faster, so they figured why not.

[@WbosonLP]

@@Nobody-sp7ug thanks for the thorough answer. I´m not arguing that figuring out some techniques might take shorter or longer periods of time. I´m just pointing out the how PR team is putting always things out as they please. Also I don´t talk about 5x time needed in GPU but this: "Trained with 5X more data than DLSS 3, DLSS 3.5 recognizes different ray-traced effects to make smarter decisions about using temporal and spatial data, and to retain high frequency information for superior-quality upscaling." 5x more data wouldn´t mean 5x more resources in $ or time, but still 2-3x more? So atm they came out with the DEMO and it´s still not perfect. But how it should work in future? Instead of making the life of devs "easier" by introducing RR they will be just charging $$ for RR data? To me it seems devs have to use resources anyway in the end...

[@Nobody-sp7ug]

@@WbosonLP Ah. I've trained myself to be prdeaf lol. So what younare talking about "trained with 5x more data" is kinda prspeak, you're sniffer is definitely smelling some bs, but its not completely bs. Always keep an eye out for words like trained. When you see that, the context has shifted off whatever the hardware or software related to playing games is, and to nvidias office and their supercomputers. So the ai model they are training on their supercomputers, which nobody plays games on, they used 5x more data to train the ray reconstruction model for dlss 3.5. This doesn't directly mean anything for any hardware games would be played on. It results in the ai model spitting out a better and perhaps larger 'guide' for the game playing hardware to follow. Nvidia COULD sell their ai models, but that would be FANTASTIC for pretty much everybody but nvidia in the long run. Nvidia having a monopoly on the by far most advanced ai models and keeping the secret recipe to themselves is definitely more lucrative for them.

[@un1ess57]

Hello. is this Demo public？How can I download it，Plz

[@olimatou9889]

You missed the sharpness slider on the demo, what does it do ?

[@danielowentech]

It increases the sharpness

[@0rdyin]

How can we get access to this demo..?

[@t3amb4sh]

I noticed very high AVG PC Latency. Why is that happening? You didn't even try FG.. Also, why can't we, consumers, download this tech demo and try it ourselves?

[@olddadagamer]

7:29 look at all the light reflections moving with path tracing off then when path teacing gets turned on, look at the light reflections grt brighter but no light reflection movement. specifically on the bars cabinets on the bottom left wide

[@olddadagamer]

* side

[@OniMirage]

I love that they made this nvidia demo for the uh .... looks better than native crowd that existed prior to ray reconstruction. lots of people gonna eat their words on this especially since this demo is best case scenario and can be flipped in real time

[@Teapose]

I like how the ceiling lights apparently aren't intended to ...light the surface below them at all, and almost everything in the scene is being lit by the strip at the base of the countertop. Great demo! Watching people say "Wow very cool" is dizzying. The marketing saturation is complete, I have no idea what people think good lighting effects is these days. Give me baked in lighting, I want my 90FPS back.

[@CeceliPS3]

But why is the illumination "crawling" to begin with? This doesn't happen on games. Is this demo made this way to force a bias towards RR? Weird...

[@hadroncollider17]

In conclusion, consoles are holding the industry back by a crippling degree

[@adriancioroianu1704]

Absolutely, but its not the console itselfs, its the way economy works. Game studios are all aiming to gain as much as possible by targetting primarely consoles, because most sales comes from there. Imagine a world where game studios would focus only on PC optimization and development. The level of fidelity and performance would be unreal, pretty much movie like.

[@Wobbothe3rd]

*AMD APUs you mean. It isn't a law of nature that consoles have to use AMD.

[@nossy232323]

Always has been.

[@pliat]

@@Wobbothe3rdno, it’s always the consoles holding everything back. Whether it’s amd or not doesn’t make a difference.

[@AranShahi]

You expect average people to buy rtx 4090. Its inexperienced dev who still dont know how to fully utilize the console resource that is holding back.

[@iwillnoteatzebugs]

Where do i download this?

[@Mayeloski]

Just got a 4070ti for cyberpunk 2077 and dam, 1440p with path tracing, RR and FG seems really well made, I hope if on next gen we got something like DLSS 4 that would be available not only for 5000 cards.

[@PhilosophicalSock]

Next big thing to release is NTC. Because of how much Nv is blamed for putting less Vram. Dlss4 is not yet even rumored

[@Z3t487]

​@@PhilosophicalSockNeural Texture Compression?

[@PhilosophicalSock]

@@Z3t487 yep. might be another game changer for Nv (if it really works as expected)

[@mariop8101]

Good luck with your 12GB VRAM and Path Tracing + RR at 1440p, 8GB not enough for a 3060ti run path Tracing + RR 1080p High/medium DLSS Balanced, It micro stutters at 40 fps.

[@tmsphere]

some ppl just get used to 30-40 fps, dunno how they do it but they do.

[@antont4974]

Well, I don't think they made it intentionally advantageous, however Performance/IQ benefits are massive when combined, tbh

[@alpha007org]

When you will be able to turn Ray/Path Tracing ON on middle range GPUs (around 500 EUR now) without massive performance hit, I really hope game developers take Ray Tracing like just another graphical option. Even if you have 4090, and you're using it for gaming, and not for making screenshots and running benchmarks, ray tracing currently add nothing to the gameplay experience. Except slash FPS from 200 to Slideshow. Then you have to DLSS upscale, turn FG to get decent playable FPS. And FG in my mind is a dead end.

[@PixelShade]

I honestly think that this generation will dabble with light RT effects, next generation will have RT effects, and the generation after that will likely allow path-tracing.... In any case. I don't think we should exclude rasterized graphics just yet due to the massive performance advantage. For me personally I would much rather want VR to take off than games to have path-tracing. The reason being that ray-tracing is fundamentally just "rendering improvements"... Meanwhile VR actually changes the way we play games and allows for depth-perception. Looking at a game like Half-Life: Alyx, It looks ray-traced... because it is... by baking ray-traced GI onto its environments... This ramen shop would also look almost equally as good with baked lighting and parallax corrected cube maps for reflective materials. The only time RT or path-tracing is required is when we require dynamic content for a scene.

[@maag78]

I wish they would give the public that demo too though.

[@Intrinsic16]

I have an RTX 3060 12 gb. Will thr dlss 3.5 work on it?

[@RIBANNI]

yeah it should

[@MANDYBLINGZ2]

10:42 I can better tell the 3 dimensional depth of the egg too, Daniel.

[@kamm3021]

Very nice

[@RubbingPotatoes]

So to summarize, DLSS < native < DLSS + ray reconstruction. Thanks for the correction.

[@Forty2de]

Not frame gen. Ray Reconstruction.

[@BlGDaddyRob]

remember when they used to show a lightbulb swinging around in a dark room to show better lighting? Now we get a pristine, static environment to show off lighting lol. The marketing Nvidia is doing with ray tracing for 3 generations now is masterful. The way they market this stuff reminds me of Skyrim modders that have a screenshot build basically that can't run for shit in motion haha. Maybe next gen I can get a card that can do solid ray tracing in more than 1 game for less than 2 mortgage payments, but Im not gonna hold my breath.

[@Blearky]

So without knowing anything particular about the direct way RTX and all the new lighting systems work I still have the feeling they are influenced by the amount of frames they can work with. Meaning not only that more FPS look smoother but also produce better results detailing stuff. It is kinda like the amount of samples the tech gets to base its calculations upon and the more frames it has the closer it get's to the perfect solution. I also disagree on the DLSS Quality vs nativ 1080p debate. Skipping between 1:21 and 1:26 I'd argue that DLSS Q. has more details than native somehow. Even the aliasing was way better when you look at the grill patterns on the left. Though I'd agree that this is also depending on the situation since the glitching on the floor patterns was really bad in DLSS. I wonder though why you haven't shown us frame gen. I understand the whole input lag debate but there is already a solution just needing to be implemented at that point if I am not wrong and as of my hypothesis I would have loved to see the results. Anyway great video as always.

[@conpa18dany]

Honey wake up! Daniel Owen just uploaded a new video!

[@vidfreak56]

Well im sure there is some sharpening happening somewhere, but DLSS tends to be pretty good. You should try to sharpen native to really bring out the differences between it and DLSS. Strange you cant do RR w/ native. ILl bet that looks even better.

[@drblitz3092]

Good Stuff, good stuff. Wish I was born in 2010, so pathtracing would be standard in my 20’s

[@Sausager]

you forgot the GPU in the description of the video

[@-Green1]

Should I get a 7900x or a 7800x3d for a mostly gaming rig? I like to get as many years as I can from my CPU's. I lean towards the extra cores and threads myself for the future but I don't know. Any thoughts from anyone is greatly appreciated. Also saw an amazing bundle at Microcenter for the 7900X.

[@jjlw2378]

Personally, I would scoop the cheap 7700x/7900x bundle from Microcenter and then upgrade to Zen5 when it eventually comes out if you feel the need to upgrade.

[@nossy232323]

Get the 7800X3D FOR SURE if you want to use it for gaming!

[@technologicalelite8076]

80% of recent-ish games (especially VR games like VR Chat and Onward if you're into that) have been treateing the extra L3 cache like water. Plus cyberpunk is jist now finally getting an update to utilize 8 cores in a CPU, step ahead most games but you'll most definitely be GPU limited before you get to that level. Games like starfield perfer Intel still but if you need/want an AM5 socket CPU (should definitely last you a few generations), 1000% get the 7800x3d

[@darkfire3691]

@@jjlw2378 lol what's the point of your upgrade? 7800x3D already has performance of regular zen5, though zen5x3D will outperform it ofc

[@2intheampm512]

Go for the 7800X3D if you're primarily gaming

[@nunosanches3693]

i think i saw some dust flying around thats why pan was freaking out

[@skylarknox9207]

You should grade our tests Mr. Owen 🤩🤪🤑🥳🦅‼️

[@AkaiSuzume]

Honestly, Give some of these technologies about 5 more years, Some time for hardware to catch up, and we're gonna have something truly special.

[@happyfarang]

ray reconstruction is not a denoiser pr say. It's AI predicted ray bouncing so in effect you get more "simulated" rays then you do with raw path tracing. But the result is less notice.

[@FOGoticus]

Curious as to why this isn't publicly available.

[@radnedge1983]

How much easier is RT for developers? I've heard it reduces a worthwhile amount of the tedium of creating game graphics, but I don't know much about this stuff. I'd feel better knowing RT has value beyond just improving graphics, because I've personally been content with game graphics for a while anyway. I mean, I played GTA5 on PC (2015 release) yesterday, and I still think it looks great.

[@tmsphere]

Pay more for fake frames so that devs have less work to do.

[@BenCH417]

it does looks better but the oil paint filter is much more noticeable with RR than just dlss

[@elfferich1212]

True it’s not perfect but still we should be happy with this. The reduced shimmering everywhere makes ray tracing so much more playable.

[@itsjustdroid]

We REALLY need to get confirmation from Nvidia that Ray Reconstruction can work without path tracing.

[@OneTwoMark]

DLSS does look better than native anyway when ray tracing isn't on. It acts as a good AA.

[@McLeonVP]

In objects always look better

[@McLeonVP]

8:01 No... RR always give better fps because RR upscales the reflection too lowering ray count (i suppose )

[@atnfn]

If a game was actually made like that tech demo a 4080 would be the minimum system requirements.

[@LilMissMurder3409]

We just have to live with some artifacting until the day comes when RT cores are numerous and powerful enough to put Monte-Carlo-style ray scattering to bed. If we can sample all pixels with enough rays for accurate lighting then denoising and RR won't be necessary and you can just rely on upscaling for extra performance and it'll look fantastic.

[@micb3rd]

Nvidia recently said the opposite, in a DF interview that the future is not just bigger and more powerful GPUs so being stupid and wasteful with rendering, the future is smart rendering using AI to be as efficient as possible.

[@LilMissMurder3409]

@@micb3rd I watched that interview; I'm not suggesting AI postprocessing will fall away, however I will assert that it's absolutely inevitable that hardware will become more powerful even if you use current efficiency levels as a baseline. We haven't hit a hardware ceiling yet and AI post-processing isn't going to create one.. AI rescaling is essentially an ingenious hack to create playable framerates and while Nvidia is justifiably proud of RR, you can bet your sunday best they wish realtime deterministic multi-bounce multi-ray per-pixel RT was a reality using hardware resources only. Having to train your algorithm on expensive supercomputer time for every game release isn't particularly efficient but they'll do it for CP2077 because it's their killer app showcase.

[@MarsMarsTitan]

I want to ask you a couple of questions.. but while playing, does one start looking at the details of the eggs.. or does he play? Is this technology really necessary to play? 🤔

[@mojojojo6292]

Dude everybody knows the quality of your egg rendering can make or break a game.

[@MarsMarsTitan]

@@mojojojo6292 🖖😁

[@guitaristkuro8898]

Still waiting for RTXDI in any normal RT game, so far I’ve only seen it in the Nvidia demos, and some UE5 videos. No reason not to have it, way better image quality, way better lighting, and way better performance (up to 2x). It's not limited to path tracing at all and has been around since before real time path tracing was announced or possibly even considered.

[@eliceomartinezmejias867]

Interesting how going up on resolution scale you only lose between 40% “performance “ (it’s not a game). 39 fps (1080p) 25 fps (1440p) 15 fps (4k)

[@lemonhaze715]

That egg 😳😳😳😳😳

[@serloinz]

have a shot every time he says 'significantly' ;p

[@carpelunam]

At this point Id much rather have higher res textures then more realistic reflections, especially since its so demanding at the moment for GPUs.

[@McLeonVP]

1:52 DLSS quality looks better than 1080p native in almost all games Except Warzone A good example is Cyberpunk Hitman WD, etc Usually is better than Antialiasing

[@snowpuddle9622]

just change the angle instead of goin "well maybe there's something behind e that's moving"

[@epicwarding]

am i the only person who wants to know why he didnt use frame generation?

[@epicwarding]

so no one else cares?

[@thanatosor]

This guy filled every component for sponsor/amazon affiliate links but forgot to mention the most important thing for this video is the GPU.

[@marsovac]

I like RR, but it still is "fake it till you make it" :D Maybe they cannot add as many RT units in harware to double the ray count so they add AI to fake denoising. It will be flawed, since is is basically a mind tought to add pastel passes over noisy images, which will provoke some artifacts and loss of proper small reflections that gets interpreted as noise since it is not a human mind but an AI one which doesn't understand the concepts of lighting at all. The good thing is that it now gets the full resolution image to denoise, so at least we get better resolution in reflections. Hope in some future we will be looking at real lighting, perhaps when we will have 4 bounces done in real computation.

[@kadoferusuraimu]

As AI rendering matures, brute force rendering of pixels will fall further and further behind in image quality and speed. DLSS and its competitors are going to improve much faster than better hardware will become available so the old way of rendering every pixel will just become more impossible and less desirable as time goes on.

[@justanotherlikeyou]

Ray reconstruction is one of the reasons I went with Nvidia over AMD. Not only better RT performance overall, but now better image quality to boot.

[@Skeames1214]

As cool as the tech demo is, I don't think it's necessarily representative of what we're gonna see in game (for a little while at least) based on the Cyberpunk PT RR update. There are aspects of it that are an obvious win (reflections being the best example, shadows are also a bit more stable) but it's not nearly as transformative in Cyberpunk as it is in this demo. And that makes sense! It's a tech demo to show off the potential of the tech. I have no doubt based on Nvidia's track record that they'll get to a great spot with RR, but right now I don't think it's quite as game changing as it was made out to be. That said, I'm sure a year or two from now most of the issues it currently has will have been sorted out. Ghosting should be the number one priority imo, Cyberpunk PT already had issues with that, and turning on RR just makes them more apparent. I still do turn it on though, lol. Those crispy reflections are nice.

[@Nobody-sp7ug]

Ghosting is more apparent with rr on? Thats a pretty big fail for whatever is going on with rt overdrive + rr, as rr's ai model was specifically trained to recognize the past pixels resulting in ghosting, and discard them. Which obviously isn't happening right lol.

[@Skeames1214]

Yes, some of the elements that already had ghosting issues are worse, and some are better, but overall it's definitely not an improvement in that respect. It needs a lot of work, but I Nvidia has enough of a track record to have some faith that it'll get there. @@Nobody-sp7ug