Inside DLSS 4 & Nvidia Machine Learning: The Bryan Catanzaro Interview

[@TalonsTech]

Left = hair reconstruction ON Right = hair reconstruction OFF

[@UlyssesM]

😂

[@THEcoolGAMER60]

thats wild

[@astralmaze]

Only on the new DLSS

[@missfortune9269]

Right? Insane that he didn't ask him about the new RTX Hair technology.

[@87Radim]

😹

[@GeneralKenobi69420]

As a bald af 29 yo guy I feel personally attacked by this video

[@TrueDesire존]

I shave down my hair every 3-4 weeks 😂

[@HeliumFreak]

You don't look that bald in your profile picture

[@TrueDesire존]

@@HeliumFreak power of AI.. 😎 jk jk

[@SayakMajumder]

@@HeliumFreak that's jedi-mind-trick

[@InFlamesor12]

you shouldnt care as there is a saying that on a dumb head the hair stays.

[@StoneyWoney]

I like this new Nvidia Hairworks version

[@soulsbourne]

Be careful there could be a hidden out of sight invisible cow with even higher quality hair rendered to handicap competition.

[@cridgit001]

Well done.

[@nomchompsky3012]

It's DLSS (Deep Learning Super Shampooing).

[@ThePaulg123]

Don't ever count out AMD's new FSR version (Flowing shampoo rays)

[@pixels_per_inch]

Actually it's called RTX Hair

[@fruitcake4910]

Bryan's hair is the ground truth for all the innovative hair techniques

[@avikde9631]

yeah i heard his hair contains all his powers so a gang of ppl are planning on stealing it

[@AJsWorld]

Comments be like: DLSS/AI ❌ Dude’s hair ✅

[@Demmrir]

Frankly I'm more interested in those fuckin' glasses.

[@oliverdavis4642]

We've come a long way from "does GTAV run better on PS3 or 360?" to technology that a lot of us don't really understand but still watch the content anyway.

[@TheEdmaster87]

@@Demmrir I wonder if those are simiilar to the ones that has been on CES 2025, where you can change the shading in real time

[@Die-Coughman]

Say what you want about the people that market and price nvidia's cards but I am so thankful for the engineers in Nvidia's R&D

[@Ojas008]

Nvidia's sales and marketing team can be the reason why people won't like current GPUs but the actual technical/ engineering team do deserve praise and the product themselves are wonderful . I wish I could get a chance to work with the team there.

[@Unordinary-lg4yt]

Of course. Luddites hate this.

[@th3orist]

its clear that the people who actually make these gpu's are some of the smartest people on the planet in their field.

[@ThunderTheBlackShadowKitty]

Frame generation bad. Super resolution good. That's just a fact, it's not up for debate.

[@Mcnooblet]

@@th3orist No doubt paid well too. Unfortunately when the average persons peanut sized brain thinks about costs, they only think about the hardware cost to manufacture vs the retail price tag, so GPUs should be much cheaper. They don't think about how much these engineers are paid per year, to put new tech into a new GPU generation, which obviously isn't "free", and they continue to be paid to continually update the tech and improve it, while being paid the whole time to do so, which in turn gets put into the cost of the GPU and one of the services you are paying for. Heck the average person didn't think they should pay for anything except raster. RT cores should be free, tensor cores should be free, and in turn collectively thought AMD had better price 2 performance ratios because it was limited to raster performance only. I think AMD was understandable to not put too much money into this stuff due to their fan base embracing lower end feature sets, until eventually realizing the fan base isn't enough alone to pay the bills and keep the lights on, and they need to capture quality seeking buyers, rather than quantity works for all but doesn't work that good, type of thinking people. To this day I see them gas lighting that all these features are the same, they should be embarrassed, but they obviously don't know any better and convince themselves that FG = FG and everything is as good as what they have currently experienced. Every version of FSR is the new "dlss killer", etc.

[@conorcost]

Fabulous hair

[@James2t3]

RTX Hair

[@letsplayblis1115]

The hair is strong. Real strong

[@djpep94]

@@James2t3DLSS upscaled hair

[@jorge69696]

Rendered with RTX Hair. He is actually bald.

[@50H3i1]

Quadrupled his hair with dlss4

[@stefanmuchel4240]

That was truly a great interview. You can tell that Alex knows what he's talking about, and it's clear that Bryan respects Alex's work and Digital Foundry. The two of them spoke as equals. It felt natural, was enjoyable, and you could learn something from it.

[@NeovanGoth]

The most enjoyable thing for me is how far away this interview is from the naive debates of people who don't understand anything beyond "bigger number better", and judge "optimization" by how well a game performs on highest settings in whatever hardware they happen to own at the time, and whatever arbitrary limitations they set.

[@PseudoPolish]

@@NeovanGoth Yeah. Typical drama queens, that are full of resentment and simply too broke to pay for good things: 1. Judge "optimization" on the experimental settings: *check* 2. Running on their RTX 3050s: *check* 3. Arbitrarily locking their own hardware they paid money for with clueless limitations (playing in strictly Native even if it looks like crap): *check*

[@eggcitedbro]

10:11 “Technology is always a function of the time in which it’s built” is a great quote from this - great engineering perspective to implement solutions with new technology based on what’s currently available. This goes double here since they’re literally the one who are paving things to work this time around by putting out tensor core consumer GPUs out there over the past few years.

[@zedsdeadbaby]

I really like his frank explanation for why they decided to drop the optical flow sensor for mfg. It took them years to get it right, and they dropped it after one gen in favor of a new tensor-heavy approach. Technology is a function of the time it was in, and sometimes you just have to move on if you want results. No wonder nvidia is on the top of their game

[@guspaz]

To be fair, Intel implemented framegen with an ML-based optical flow engine from the start with Battlemage. They're also doing it on much less capable GPUs than high-end 3000-series cards, so the decision to do framegen on RTX 3000 really is at this point a business decision, do they put in the time to make it work or not.

[@durden0]

@@zedsdeadbaby Even wilder is the fact that, probably within 10 years all this tech will be be obsolete as games will be entirely AI creations with the entire worlds generated in real time as you're playing them.

[@lennartmook9182]

@@durden0 maybe bu tno one will play that garbo shit

[@durden0]

@lennartmook9182 you won't be able to tell the difference by the times it matures. Just like you probably won't be able to tell the difference between native and full 4x DLSS Frame gen.

[@RafiYagami]

@guspaz Maybe modders can do it for them if they choose not to do it

[@amnforge]

Bryan's hair is DLSS version 9999; Fantastic hair.

[@Beringer90]

Looks a little dry IMO

[@ThePaulg123]

​@@Beringer90 might be a shader problem on your GPU or your driver is not up to date

[@MDXZFR]

@@Beringer90 should use reshade preset

[@maksterone4059]

I think so😂

[@maksterone4059]

Reinstallations or refrash need

[@Benefits]

To all the hardworking engineers in the tech field out there - Thank you for what you do.

Except for GenAI - shame on you forever.

[@joross8]

Not many groups who could put on an interview like this. Kudos to Alex and Digital Foundry! Great interview.

[@missfortune9269]

Bryan Catanzaro, Mark Cerny... all these industry legends watch DF and love it, that's awesome to see!

[@what-di3vb]

thank you to whoever setup this interview, and thanks to alex for having an amazing set of questions! this is a fantastic video.

[@zachb1706]

You definitely need to test the performance hit of DLSS 4 on 4000 vs 5000 at the same base resolution.

[@juanmanuel2659]

And also on series 3000 and 2000. The transformer model requires 4 times more compute, I hope the latency of DLSS doesn't become a huge issue on the older cards.

[@DanKaschel]

​​@@BucksterMcgeeNope. The context was talking about how long it took DLSS to run. DLSS3 runs in about 0.5ms on the 4090. his response was that DLSS4 does 4x the compute, implying that 0.5ms < DLSS4 < 2ms.

[@BucksterMcgee]

@DanKaschel 4:48 he explicitly explains that it *doesn't* take 4x to execute on your GPU, especially with 50 series tensor cores. 4x is the compute it took to create the model.

[@DanKaschel]

@BucksterMcgee you're misunderstanding. He's saying "don't worry, just because it requires 4x the compute, DLSS performance cost didn't just quadruple because it's more efficient and Blackwell has more compute than ada." Note that it wouldn't make sense for him to talk about Blackwell's compute capabilities if he were talking about model training.

[@BucksterMcgee]

​​​​​​@@DanKaschel I'm sorry but you are the one who misunderstood. This has already been explained and clarified by other Nvidia engineers. The 4x compute (and 2x parameter) increase is referring to the model generation. They are saying the model is far more advanced because it includes twice the parameters of the CNN version and four times the compute to train the module. But executing (inferencing) the model on your GPU is not directly portional to the model training compute requiments. Both the wording at 4:49 and 6:46 "the model has much more compute in it" makes no sense in referring to inference/execution, it's talking about the compute used to train/generate the model. Other Nvidia engineers have explained that the inference/execution cost is no more than 5% vs the CNN model. The reason Bryan is sticking to specific information, e.g. 4x compute increase for the model, is that is what the marketing is currently allowing during the embargo period. You can see him at multiple times during the interview return to these specific examples that were used during the keynote. It's the only information they are allowed to provide until after the embargoes. The Nvidia employees explaining the 5% execution drop said they cannot explain more while the NDA for the embargo is active.

[@tomd.2982]

Digital Foundry is my favorite channel... always something new and relevant every time

[@DubElementMusic]

and new sponsored content

[@itsuadman]

​​@@DubElementMusicQQ

[@Wobbothe3rd]

Catanzaro is brilliant. From CUDNN to this. What a time to be alive!

[@siriansight]

Difficult to express how much I LOVED this interview Bryan's eloquent insights Alex's customer-voice Noone ever interrupted the other, just delivered freely Amazing. Thanks

[@johnc8327]

DLSS quality mode already looked really good. Excited for this new model.

[@DBTHEPLUG]

I'm already exclusively using performance mode at 4K, so this will be a welcoming upgrade.

[@damara2268]

@@DBTHEPLUG from what has been shown, it seems that the new transformer model in performance mode has similar amount of detail as current DLSS quality.

[@Blazemaster9]

Will it lag on 40 series?

[@parthibanspace]

​@@damara2268if that's the case, it would be phenomenal. Looking forward to it.

[@Peylix]

I'm super excited to jump in and try DLSS 4 Quality in Alan Wake 2 once it drops.

[@sapphyrus]

I love how Bryan reveals to us that we have been practically faking 3D graphics for 50 years and proposes that AI generation can help with that faking being more accurate.

[@KarlDag]

I don't think it's any revelation to DF regulars, or people that bothered to look into how 3D graphics are done.

[@Morimea]

Everything humans do - is about compression of information.

[@Alberto-xw8vx]

​@@Morimea except when you experience "god", in that situation there is no compression as there is no language.

[@pixelverdicts]

That is absolutely true, AI is more capable of finding patterns more accurately given the examples. Supervised Learning is king because we have so much of data.

[@CrAzYpotpie]

​@@Alberto-xw8vxWhat? Are we talking reality or fairy tales? It is time to grow up.

[@ClassicGamingPortal]

The more they charge you the more luscious his hair will get

[@saiyaman9000]

I will pay double the 5090 for that

[@ZachAttack2U]

Shut up and take my money!!!

[@zaxmaxlax]

the more you buy the more you save

[@TEk0n]

I love these kind of “from the source” interviews always insightful keep it up!

[@EmblemParade]

Moving towards "top down" rendering is an important way to think about 3D. The "bottom" can provide the basic structure of a scene with very few polygons and "hints" about materials and lights, while the "top" would know how to fill it in based on real world image learning. It's essentially video generation. Of course we're very far from being able to generate video with AI at 60 FPS. But there might be some middle approaches we can start using.

[@shawnmacfarland3828]

Excellent interview interview Alex! Thank you. Thank you Bryan for sharing thoughts and insights.

[@76sagman]

That guy a is everything I expected him to be as the AI man.

[@matt01155]

Everything and more. Nvidia hit a home run again. I could see why Jensen is seen as a rock star in The Valley.

[@jorgeacosta6698]

Nvidia is waaay ahead of everyone else is not even funny. Say what you want about their prices, but the technology behind them is freaking amazing. DLSS is black magic

[@ThunderTheBlackShadowKitty]

This is just objectively correct and irrefutable. DLSS upscaler cannot be beat.

[@DubElementMusic]

its the reason why we have blurry unoptimized games with poor image quality, from all this games so far, very few of them implemented all the fancy new technologys good. in the past, i had good image quality, now almost all games look noisy, unstable, run bad and with fake frames. when i first saw RTI in control, i thought my GPU was broken, i could not belive that nobody saw this, everything was noisy, and smearing, and all the people talked about a good looking game. its crazy imo.

[@TheRealDJ-NEO]

@@DubElementMusichow is that nvidia’s fault? Did you try playing games from devs who actually care?

[@G0A7]

​​@@DubElementMusic so developers make a shitty game = Nvidias fault? Idk chief im pretty sure AMD have GPUs that can run those games too

[@BlackParade01]

​@@DubElementMusic The problem with Nvidia tech, is that it's all designed for 4K and for high framerates. Ray tracing will look worse at 1080p. DLSS will look worse at 1080p. DLSS will even look worse at 30fps than it does at 120fps, because it's a temporal solution and thus needs more frames.

[@twelvegrain1111]

The idea that you can train a neural renderer on real world data and offline sims to generate neural textures is truly mind blowing. These engineers are so outside the box. It's exciting to hear about the future of this technology. Really sparks the imagination in terms of what's possible and how this will augment game dev. Thanks for this video!

[@NeovanGoth]

It's pretty obvious though in some sense. Photogrammetry is already widely used to derive textured 3D models from a bunch of photos, but for modern PBR pipelines you'd also want to get more advanced material properties. So what you basically do is to use an AI not only to extract diffuse textures, but to create a whole shader program that approximates the real material's properties. There have been a lot of fascinating papers about that topic in recent years.

[@pixelverdicts]

Absolutely perfect interview!! (And wow, Alex has a torso and legs). I can always listen to DF crew. And Bryan is someone I look upto as a Deep Learning Professional.

[@Stef3m]

So now we can ray trace Alex I guess

[@pixelverdicts]

@@Stef3m Yes, do we want to do a software ray tracing or hardware ray tracing? I'll send you a design document in a short while. Also, bring 4-5 torches, we will shoot rays manually, We want realism as much as possible.

[@AiMusicPuppy]

We all love this interview! Do these more.... twice a year please ....specifically 🙏 with this technology genius ...If he has the time.

[@DETERMINOLOGY]

GLAD that DF made a video. Also for those that has tested the 50 series hasnt really had any major issues with it. But yet fans still rip on it Once these gpus hit i think people wil enjoy them more then people think

[@Jean-jk4zv]

So pleasing to watch both of you gentlemen discussing about new graphics technologies ❤

[@P4shaPlays]

25:19 I absolutely love this whole segment and especially the quote about how when painters paint a snowflake, they are not trying to physically model how a snowflake looks, they just know how it looks. A lot of people cling on to silly talking points about "fake frames" and "not real native resolution", but they seem to forget that there is nothing inherently "real" about rasterization and how we model the game world and then present it on the screen to the user. So, if we can do it faster and more efficiently with the help of AI algorithms, then why shouldn't we do it?

[@kazioo2]

All the AI generated images and videos did just that with diffusion method. When you tell Sora to make a water flow it doesn't calculate a single particle. it's just guesses how it should look the same way our brains do it in dreams.

[@jayblaze53]

Did he imply multi frame gen could come to older GPUs?

[@steinkoloss7320]

Technically yeah, but it won't happen due to Nvidia needing you to buy their new GPUs and because it's honestly going to be real hard.

[@jayblaze53]

@steinkoloss7320 It shouldn't be too hard really, especially on the 40 series GPUs as it's now using AI instead of the OFA dedicated hardware unit.

[@gosti500]

he absolutely did

[@steinkoloss7320]

@ Well Blackwell is a LOT faster in AI than Ada Lovelace.

[@gosti500]

@@steinkoloss7320 but is a 5070 faster in AI than a 4090?

[@mamuf]

Cool tech, great interview! Now I can't wait for all the actual hands-on testing you guys will be doing with DLSS 4 when it comes out. Specifically the benefits it will bring to older RTX GPUs.

[@ThunderTheBlackShadowKitty]

Even the lowly 2060 and 3050 will be able to use the new DLSS 4 upscaler. This is very pro consumer. Just remember what DLSS 1 looked like and realize where it is now. Massive improvements.

[@jorge69696]

@@ThunderTheBlackShadowKitty Just because they can use it doesn't mean it will be fast enough to actually be worth it. Pascal also had ray tracing, but it ran at like 5 fps.

[@ThunderTheBlackShadowKitty]

@@jorge69696 Nonsense. DLSS 4 upscaling is a blessing to older RTX cards. FG bad, upscaler good.

[@mamuf]

@@ThunderTheBlackShadowKitty Yeah, I'm looking forward to actual tests of picture quality and performance on such hardware.

[@ThunderTheBlackShadowKitty]

@@mamuf Image quality is always the #1 priority with upscalers, no way around it. If you can't get pristine image quality then using any upscaler is a waste of time.

[@shzmeteor3954]

This was a very interesting and informative video. I enjoyed it a lot. thanks to all involved.

[@what-di3vb]

^^^ agreed fully. these videos are informative insights on engineering complexities and NVIDIA as a company aside, it’s amazing to see stuff like this. i love that digital foundry exists!

[@Battler624]

that was a non-answer on the 3000 series question. I love the surprise.

[@czbrat]

He didn't give us any information that's new, it's all been available on their website. Really pointless interview.

[@davidsentanu7836]

​​@@czbratIt's just to show his fabulous hair. Can't read that on their website. 😂

[@rezaimran98]

I mean how else will they keep scamming people to buy new stuff.

[@deusfaux]

@@czbrat imagine ppl consume information in different mediums. just, imagine.

[@czbrat]

@@deusfaux This very channel has broken down and explained all this information already in a DF direct special. There is really no point to this interview.

[@TrevorSullivan]

NVIDIA is absolutely dominating the GPU and AI markets. There's simply no competition.

[@ThunderTheBlackShadowKitty]

@@TrevorSullivan As an AMD fan I gotta say, the upscaling from DLSS simply cannot be beat. I cannot wait to test DLSS 4 quality on all my games. I already thought DLSS 2 looked good but this...Wow.

[@xAfroMetalHead1990x]

That’s not a good thing

[@WillAmbiance]

I will say with Nvidia dropping the ball with low Raster improvements this Gen, hopefully that gives AMD more time to improve and compete more. 9070XT is looking real impressive nearing them for so cheap expectedly. I say this as an Nvidia user. If AMD has a 9090, I feel with 5090 lacking, it’d prob nearly meet it. Just my thoughts!

[@DubElementMusic]

because the industry does everything they want, still most of the ai tech suck, the reason why we have unoptimized, blurry, shimmering and flickering games.

[@RawmanFilm]

This gen? They did so for multiple. Pascal was the goat, 2000 should have been tsmc, 3000 looked good because 2000 was not good, 4000 had price to performance regressiona in raster. talking mid to low end here btw. High end was jumping leaps, although it is increasing in price each gen

[@doctorsam5771]

Can't believe people are actually complaining about this tech lmao, This is as groundbreaking of a tech as it gets for PC gamers. You can literally strong arm games into running well with this tech the way it's improving

[@GamingKing545]

yeah but you can tell they are fake frames

[@doctorsam5771]

@GamingKing545 No you can't, there is literally a video on this channel about DLSS 3 which showcases fake frames vs real frames and no one can tell a difference and that was an older model. PC gamers are just obnoxious geeks who can't enjoy anything. I remember them crying about fake resolution 6 years ago and now that the tech has progressed to a point fake resolution looks better than their native they have to lach onto a new target to complain about.

[@GamingKing545]

@@doctorsam5771 i can tell on my own hardware its obvious to me and im running a 4090 paired with a 4K display

[@GamingKing545]

its obvious to ME that its generated im running a 4090 paired with a 4K display also its not healthy to shill for youtubers.

[@doctorsam5771]

@@GamingKing545 😂

[@Peanut-Butter-Banana-Rama]

There is no ghosting or shimmering on Brian's hair and there is clearly Ray Reconstruction. In fact, we might be looking at the next version of DLSS.

[@cesarzooba9331]

Glad to see NVIDIA Hairworks is back at it

[@YoureBreathtaking]

13:06 that answer seems vague, I guess they're gate keeping the new tech under a paywall. So RTX 4000 could use multi frame generation and RTX 3000 could use frame generation since it only relies on tensor cores and not optical flow accelerator, it just needs software update?

[@Monsux]

These DLSS 4 hair physics are nuts.

[@ReedGarzone]

I have a lot of respect for you, Digital Foundry. But I’ve lost some now. I might have expected this from Oliver or Rich, maybe even John with his generally close cropped style. But you Alex? The most harry man on Digital Foundry and not ONE single question about this man’s hair care regimen.

[@ghost085]

Where do you leave Tom and his fabulous hair?

[@dam.s]

18:40 Love seeing The Finals in this video. Great modern, fast game to showcase this technology-would be neat if used in a more in-depth video on the topic.

[@Crimsongz]

It’s the best optimized UE5 game I have ever played !

[@ArathirCz]

Another great interview with Bryan. Thanks Alex and DF team.

[@360Fov]

I appreciate nvidia for their staff interviews....even going back to Jansen's interview/presentation at Harvard back in 2013, they seem down to Earth and super passionate about what they do. Hail DF for focusing on this ultra-technical aspect of computers and gaming too!

[@Stef3m]

A bit like Elon

[@ThunderTheBlackShadowKitty]

I don't think most people understand just how important this is. Image quality is the #1 concern when upscaling. These improvements are massive. This makes upscaling from 720p to 1080p or 1440p viable.

[@BlindBison]

For the section discussing traditional non-VRR V-Sync you are heavily underestimating the number of people for whom that scenario applies. A ton of users do not have VRR displays even in 2025. What Intel did with Smart Sync for example is very clever and I still think traditional vsync innovations have value for many users and use cases.

[@DemonGamerTT]

Are you guys planning on doing a second video on Lossless Scaling 3.0? The new version is out there and has improved exponentially, its very hot topic ATM.

[@davemeredith6964]

Very much doubt it...pretty sure Nvidia put a nice sponsor on this??

[@TTx04xCOBRA]

@@davemeredith6964yall are so dumb. They did a big video on FSR 4

[@CFTStheband]

I could see DF doing a new video if there was a marked improvement in artifacts and stuff. They obviously seem really sensitive to them 😂 I think I tried LS a few months ago in v2? It wasn't worth it for me personally and the games I play, but if there have been good quality updates it would be interesting if they did because a lot of people seem to find it useful.

[@damara2268]

@@CFTStheband the v3 has MUCH less artifacts than the previous ones.

[@CFTStheband]

@damara2268 I believe I did see recently that the framegen goes up to 20x now in v3?! 🤣 If you happen to know of a good video analyzing performance and taking a look at improvements to artifacting as well as comparison to Nvidia and AMD framegen, please drop it because I'd love to see it. I gotta try it again myself one of these days in any case.

[@plume...]

Thanks Bryan and Alex, great interview

[@theata]

To the engineers working on this: You guys are making the next generation possible. I know a lot of people are harking on AI and fake frames, but the reality is, the cost of quickly rendering improved RT scenes is just not something modern consumer cards are capable of doing independently. And while many will say, less RT more raster, the reality is the cost in time and development to create those illusions is part of what's making development cycles move like molasses. I still think the tech is in its infancy->moving into to toddler era, but it's clear having ray traced lighting is kinda where we want to go. And unless some miracle in hardware suddenly poofs into existence, the need to use AI in artificially generating frames in between the real ones is going to take center stage. So I look forward to the tear down and independent metrics, but keep up the good work guys. I'm curious what language they use to train these models. Is it just a pytorch set up, or are they using something like C++ to optimize memory and performance during each epoch?

[@Mjet403]

I play almost all titles with DlSS Quality mode very rarely can i tell the difference especially in motion plus who like high frames 🤘🤘

[@Ferdinand208]

Now we need a new John Carmack that creates an engine that just runs on the shaders,tensor, RT cores and no longer uses the fixed function parts.

[@Aetlaify]

He was a god among mortals when I was a teenager.

[@pixelperfectpizza]

He's working on agi now lol

[@kazioo2]

Fun fact: the warp method in Reflex 2 is partially based on John Carmack's research in an article "Latency mitigation strategies" from 2012.

[@Ferdinand208]

@@kazioo2 that is so cool. It was probably impractical to do until you have matrix calculations and enough software development. DLSS4 is also based on the same research as chatgpt.

[@NeovanGoth]

⁠@@Ferdinand208There have been various demos over time, which worked pretty well in terms of reduction of perceived latency, and only failed at the inpainting of missing parts, which is the much harder challenge.

[@dogussahin]

So much packed information for a 28 minutes interview, well done

[@remij_9487]

Bryan is awesome! Just incredible work from him and the team. Great interview Alex!

[@Nnamz]

Stellar interview, Alex.

[@donalgodon]

I would love to see a detailed breakdown of what the new features mean to the 40 series cards.

[@mike_miller]

yeah like how much FPS should i prepare to lose compared to CNN

[@msnehamukherjee]

​ @mike_miller I doubt Nvidia would release transformer model for older GPUs if it's really that performance heavy. It will basically negate the idea of DLSS and why we use it in the first place. The performance overhead is probably going to be minimal at most.

[@g60force]

@@msnehamukherjee he said FG 2x 3x 4x will be selectable via NVapp cuz they don't want to rely on DEV adding options, so it would be safe to assume 4xxx owner could use NVapp to choose between CNN or Transformer!

[@LuizFernando-li1du]

@@msnehamukherjee Maybe it will have a significant cost, but you will be able to use more aggressive presets (Balanced, Performance) to offset this cost and still get better image quality.

[@Khaostheory1980]

Really great job DF, you are the masters at this.

[@Netsuko]

Kinda crazy that a 4 year old game is still THE go to when it come to graphics. Goes to show just how good this game looked even in 2020. The RED Engine is just insane. I understand the why, but I am still sad they are moving to UE5. Everything is unreal engine these days.

[@User9681e]

The savings in development time , training for their engine Should allow better games to be developed

[@riven4121]

@@User9681e More like REDEngine being problematic to work with. Half of the issues with CP2077 was because its engine outgrew CDPR's ambitions, the other half was CDPR lying.

[@PortfolioPL]

Most AAA games will keep their own engine just as before.

[@CrAzYpotpie]

​@@PortfolioPLNo they won't. If you do any research on that, you can see many long standing in house engines being dropped by huge companies in favor of UE.

[@PortfolioPL]

@@CrAzYpotpie A little more than usual with previous UE versions but still, most AAA games will keep their propriety engine as always. Huge risk CDPR took with this and most of us are better off with out own engines we control fully. And I worked on AAA games in UE 2-3-4.Although dealing with Epic is good now while they want us to switch but in a few years time they will start forcing higher fees and devs who switched like CDPR will not have a choice but to pay whatever Epic Megagames demands. They are already doing this with movie folks...

[@jjdizz1l]

This man is the future of 3D graphics.

[@yousuff1]

I'm really hoping the new DLSS lets you go one step down while maintaining the same quality as previous gen. ie. DLSS 3.5 Quality looks the same as DLSS 4 Balanced. That would really help the 2000 and 3000 series last a while longer.

[@ThunderTheBlackShadowKitty]

Agreed. I want to get to a point where I can't even tell DLSS is on.

[@nicholasmalaventura5710]

it's strange bcs nvidia siad that at ces but in this interview at 4:40 the interviewer said the opposite

[@xTurtleOW]

You can manually change dlss version from game files or just dlss swapper application

[@kev1nno]

@@nicholasmalaventura5710 IIRC what Nvidia said at CES was that the new version of *Frame Gen* would both look better and run faster on 4000 series GPUs than the old version - not super resolution. Although I can see it being easy to mix up considering Nvidia call both frame gen and super resolution "DLSS"

[@pondwhale]

it wont be an apples to apples comparison, because increasing the quality of SR didnt necessarily help with issues like ghosting or moire patterns, as this new model apparently does. a new model also might mean new pain points that we wont know about until release. my expectations are pretty high though

[@killerfugu2659]

Bryan is a pleasure to listen to and great at explaining, content like this when new tech drops would be amazing.

[@PishogueEsq]

You can’t fool me. That’s Keith Raniere, cult leader of NXIVM!

[@urazoktay7940]

Amazing video, thank you Digital Foundry. Very interesting and intriguing topics.

[@iROHxo]

15:37 V-Sync is important for frame generation mode as well because if you're connecting your TV (most TVs don't have V-Sync) to a PC, you simply can't use frame generation. Not to mention, almost all TVs are 4K, and if frame generation is needed anywhere, it's on TVs.

[@11cat123]

And most TVs are 60hz as well. The few that aren't will have VRR support. Frame gen has never been a 30fps to 60fps solution. It has always been a solution for saturating high refresh rate monitors with frames that would otherwise require too many sacrifices in settings and/or resolution to attain considering you are already at 60fps.

[@zachb1706]

There are Gsync compatible TVs though aren’t there?

[@iROHxo]

@ yeah you are right if we talk about 1x frame gen, but if we consider 4x frame gen of this current gen tech it might be solution to 25-35fps raw to 60fps as well.

[@iROHxo]

@ yes but only high end. some tv's even have it but manufacturers are not enabling it, like LG for example they have A series which is basically identical to B and C series Oled Tv's except A series has no VRR enabled even though Chip is capable. most Samsung TV's doesn't have VRR as well.

[@zachb1706]

@@iROHxo well as the guy above pointed out, frame gen isn’t really made for low end TVs. You need atleast 120Hz, and a decent response time to not feel awful

[@forthcomingphenomenon]

Thank you Bryan and Alex! Great interview!

[@zyxyuv1650]

Why can't NVidia's Frame Warp predictively increase the resolution+viewport size to over-render past the edges of the visible frame, instead of filling in black holes with a generative model? Even if you enabled over-rendering permanently it could be a 5% performance hit to render a larger frame across all the edges, but the prediction could be much more efficient than that, and it could fall back to a generative model only in the worst case of misprediction.

[@adamdincorn]

I've thought about this idea of over-rendering as a potential way to mitigate screen space reflections always being cut off at the edges. There doesn't seem to be a framework for it yet, but hopefully some kind of solution could be made

[@videogaminbiker889]

could also over render in a lower resolution

[@Hyperus]

Over-rendering only works if you only modify rotation, but position is also modified here, hence you need to infill behind objects anyway.

[@Hyperus]

@@adamdincorn This is a well known technique and game implementation wise it is really easy to implement. Blender even has this as an option called "Overscan", but its not exactly cheap. Adding some 10% on all sides would render 21% more screen estate, resulting in visibly lower performance. If I made a game/engine I would certainly include it as an option though.

[@EDcaseNO]

on the topic of reflex (reducing input times), that would be a heavy tax.

[@san.cochado]

Beautifully done!

[@PhilomenaMerritt-e5e]

Thanks for making this video! It's so helpful and interesting.

[@shadowscreed17021]

This is a very good discussion. The future is not so far away. A.i. will be helping out understanding the real world in design. Lights, shadows, Ray tracing, rain, and more. it's all up to what the artist wants you to see and be affected by while taking in what dialog there is.

[@Bry.89]

Should've asked if Reflex 2's time warp thing works with Frame Gen

[@BucksterMcgee]

It sounds like it doesn't currently, but the goal in the future is to combine them, which would seemingly eliminate the latency cost of frame generation (at the cost of more artifacts).

[@foghornblue]

Find someone who looks at you the way Alex looks at Bryan Catanzaro

[@paulbrooks4395]

I noticed that he couched his words around reliance on Tensor cores and Tensor compute as it fits into the time-to-render budget. We may be able to extrapolate that older cards below a certain spec threshold of a given generation may be incapable of running the newer models in a timely fashion. It's possible they could work at lower resolutions on older gen tech, yet I would contend that provides less experiential benefit due to the lower input resolution/data that comprises a scene. In short, probably only a number models of 4th gen and maybe 1 or 2 from 3rd gen. The other half is that they can take advantage of newer algorithms that didn't exist prior to our greater experience with AI, and secondarily, Nvidia is incidentally now forcing future graphics to rely on hybrid rendering as a consequence of Blackwell being marketed to AI companies. Technology is a product of three items: diversity (Tensor and RT), complexity (more systems and subsystems interacting), and ubiquity (universal acceptance of a new paradigm--AI-in-the-render-pipeline as a de facto standard). I foresee that Blackwell will incidentally become the new foundation series of GPUs that include massive amounts of Tensor capability.

[@infinitysynthesis]

Nvidia's version of Mark Ceny.

[@faceurhell]

Nvidia has far, far, far more advanced, powerful and revolutionary tech than PlayStation.

[@laserbeam3836]

Mark Ceny wishes he have this man talent

[@infinitysynthesis]

@ lol good job taking my joke and being all serious with it 😆

[@pedro.alcatra]

Hell no

[@joeykeilholz925]

​@@faceurhelllamest shit you could have said. Congrats!

[@NathanYoungman]

Thanks for the interview. It would be interesting to learn more about the neural rendering features in the 50 series (texture compression, etc.)

[@dolan_plz]

Frame pacing was something that was discouraging me from using frame gen the most. Not even latency, but frame pacing was too erratic for my liking. Glad they are announcing improvements in both these areas! 🙂

[@lynackhilou4865]

What's funny is it seems fsr3 doesn't have this issue , but it has a little more artifacts and sometimes issue with UI . Frame gen from both seems to have a lot of caveats to use

[@LadrimSlump]

I hope that we see some improvements on our series 40xx... FG framepacing killed me

[@sh_chef92]

??? Just use gsync, vsync, lowlatencymode ultra / reflex on together and if you are not hitting that automatic fps cap that is few fps under your refreshrate, than lock your fps manualy slightly under your average fps. Problem solved.

[@LadrimSlump]

@@sh_chef92 we are talking about Frame generation, even with all the above on, framepacing was awful

[@pixels_per_inch]

Same here. I can't use it without feeling like I'm going to vomit.

[@AdityaTripathi]

This Neural Hair engine is insane, well done!

[@Zeeflyboy]

I would love someone to ask about if they have any intentions to bring some of their more interesting tech to VR - where we could really do with more performance. Of course we have DLSS super res in VR too but a driver level replacement for Asynchronous reprojection (ASW/SSW/Motion Smoothing) using the depth and motion vector data that DLSS already has access to would be a huge step forward... would love to see Nvidia throw VR a bone again after a long hiatus.

[@nrosko]

I wish someone would at least ask this question. I guess they will just say its not worth their time. Even Meta seem to have given up on PCVR despite many linking their quest.

[@doyouthinkitsdead]

Dude thats fascinating stuff nice one!

[@DragonOfTheMortalKombat]

Can't wait for Reflex 2 and DLSS 4 transformer model

[@rangerpopo6702]

My problem with frame generation isn't that it uses AI or "guessing" a frame It's that it doesn't help making the NEXT frame only between 2 frames the GPU already calculated traditionaly, so it will always only add to the latency and it means diminishing returns because how many frames can you add between 2, we need AI to help us reach the next frame as soon as possible not make us wait even more for it

[@JamieWoods00746]

Transformers more than meets the eye!! Autobots!! Transform and roll out!!!!!!!!

[@L33TBEANS]

Great interview, nice to see a more genuine glance into what problems nvidia is trying to solve.

[@Thuner67776]

So would the new model be at all useful to the older cards like 4000 series? No point running DLSS if it costs as much as running native.

[@iurigrang]

Well, if it's taking the 4090 0.5ms to run the older model, even if it was 4x more expensive (which they claimed it isn't), 2ms is still a lot faster than the difference between quality and native on almost every game you'd consider turning DLSS on. I'm more worried about older cards, like the 20 series, especially on the lower end. There I do get worried it might often be worth it to stick with the newest CNN model. I wish they would work on a lighter transformer model for both the switch 2 and older cards, and though there are leaks of a lighter model for switch, I'm not holding my breath that it will be more useful to older cards than the CNN model.

[@Wobbothe3rd]

The rtx40 series Tensor cores are no joke. Look up the AI TOPS, unless DLSS4 is totally dependent on FP4 (which I doubt), the 40 series should be fine.

[@Jonatan606]

The Nvidia app will let you choose between the DLSS models and from what I have seen from the Cyberpunk preview, you can also switch models within the game menu.

[@ThunderTheBlackShadowKitty]

@@Thuner67776 DLSS 4 upscaler will work with 20 & 30 series as well. Just forget FG exists, it's a scam technology. The upscaler is magic, while FG is garbage.

[@existentialselkath1264]

I think the new model would be useful for DLAA. Getting the best possible image quality can make a big difference, and it'd still be much cheaper than supersampling. For dlss though, you may be better off going one step higher in quality mode than using the new model

[@imAgentR]

Async reprojection (not necessarily just the inpainting + frame warp) is the way. Fully decoupling the input to match a 1000Hz monitor would be phenomenal. Can’t see atm how any card in any game with frame gen would produce that many frames. Or even want it to.

[@MartijnSpits]

I thought about this this week. This would be so awesome

[@jinmiseru]

Oh, looks like Mr Catanzaro got ahold of Dragon Age Veilguard's strand hair tech for himself. I'm jealous.

[@ThunderTheBlackShadowKitty]

Such a good game.

[@ZidaneSAID]

Thanks DF for the great content!

[@xxmujinscxx3516]

The frame time spikes is not from the developer end, it’s from the unreal engine end( industry standard)

[@tenchotenchev5606]

Yes and no. Everyone was lied about that Unreal Engine is ready to be used tool by newbie devs with no actual coding experience and this turned out to be lie. UE is free but not by any means finished product. You still got to tweak and add stuff to it like your own AA solution, and todays devs prefer to skip this entirely and take the easy way.

[@faceurhell]

Unreal Engine is a lot harder to develop for than people think, including devs when they first dive in.

[@DBTHEPLUG]

Okay, Mr. Expert.

[@JPWestmas]

This explaination of Reflex 2 does sound like a game changer. Hope it works!

[@P43YM]

Reflex 2... ok inpaint for holes when moving camera, but what about dynamic objects in the scene? What if enemy change direction in the middle of the process?

[@kev1nno]

It would probably "mispredict" in that case. Although you'd probably want to use reflex 2 most in e-sports games where you're running at several hundreds of fps, so the misprediction from dynamic object movement probably won't be that significant, since the frametimes are already so low

[@Ashnal]

That's not how that works. During that process, the game world data has not been updated. All it is doing is adjusting perceived camera position based on reading input, without updating the game world simulation. The enemy position cannot change during this process.

[@P43YM]

@@Ashnal so latency for player movement will be improved, but world latency is not?

[@P43YM]

@@kev1nnoYeah , probably you're right. It's just question out of curiosity. Maybe DF can test it somehow in the future

[@faultier1158]

What it does is reduce the camera latency, and literally nothing else. Game actions, enemy movement - everything else is at the original latency. Having more responsive camera movement is very nice though

[@ryan.crosby]

18:30 So Reflex 2 basically works the same way as VR re-projection technology (aka "Async Spacewarp" or similar) where the image is warped to match the latest known user position before presentation. I wonder if the advancements on NVIDIA's end could be fed back into the VR side and net better perceived latency in VR.

[@danielsilvasich]

Great interview Alex and great questions as well 👍

[@giantbellend]

I always look forward to hearing Alex close the videos with Auf Wiedersehen. A joy to listen to.

[@kukuricapica]

It makes sense to generate scene in time where Moore's law doesn't apply anymore. I have seen a lot of people on youtube that makes fun of 50series, because it's typical raster performance leap is probably smaller compared to 40series, but every major game title coming out will probably use RT and some of them will most likely be even using Full RT (Remedy, Guerilla games, Massive Entertainment, CDPR..) and in those titles using DLLS and having 2x faster RT performance may turn out in significant difference in performance. Keep in mind that Nvidia's engineers is talking with developers behind the scenes and I dont think they would bet heavy on all this Neural silicon architecture if they didn't believe it is the only way forward.

[@saiyaman9000]

I find it interesting how pretty much nobody has had hands on with this new technology and people are saying "more fake frames bad"

[@panjak323]

nGreedia killed Moore's law, by selling extremely cut down mid range GPUs for ridiculous price. 50 Gen only makes it worse.

[@kukuricapica]

@ You talk those engineers are some evil cabal. Problem is that we are pushing the limits of lithography when trying to cramp as many transistor as we can on a silicon chip that is getting bigger and harder to cool. And it wasn't Nvidia but AMD if I remember correctly that said it first. I don't think you really believe that we can make every 2 years a chip that has 2x performance of previous generation simply by printing more transistors. The biggest performance leap are made by design and architecture. AMD just made CPU chips with bigger cash and what do you know, Intel is WAY behind now in gaming performance. I know I will be getting 5080 because performance is great for the price in my opinion but most of all I do think that DLLS4 will get better since it's algorithmic solution and it will run the best on 50series especially in those games with Full RT where you can't get 60fps in 4K without DLLS. Anyway I think you will need to wait for some killer title that will take advantage of all these new features like AW2, CP2077, Star Wars: Outlaws and/or future titles like Control 2.

[@GeTech1]

Informative video love it❤️

[@retalaonteany7899]

DLSS > FSR > XeSS

[@MakubexGB]

The 5090 Founders Edition comes with a signed bottle of hair conditioner

[@ZomgZomg007]

So they said for FG, they dropped the optical side of it in favour of AI, which for example the 4090 has more of it compared to 5070. I still feel like the 40 series could have benefited from the MFG and the updated model of FG. Something about them giving a reason why 40 series can't have this seems super odd, and especially because both 40/50 is basically the same construct (this is why we are not seeing much of a uplift in raw performance)

[@Die-Coughman]

You're probably somewhat right but remember core count is only half of the picture. The Tensor cores have gotten new features that may allow them to run the models more efficiently

[@pixels_per_inch]

They're just gatekeeping to sell more 50 series, otherwise the performance they showed off wouldn't be that much better than a 40 series

[@Andytlp]

Something like tensor cores, theres no way to tell if say 4000 series 4090 300 cores or something like that are actually faster than say 150 cores on a cut down 5000 series gpu or about matched. If they say performance is doubled then you just assume it is. Its much more simple with raster cores. Faster clocks, more memory bandwidth you can calculate the speed boost more or less within 10-20%. Rt cores or tensor needs official numbers to know. Or when benchmarks come out. 5090 is like 30% faster in rt, dont know how much better tensor is, 2x?

[@GiuseppeBeatz]

So stoked that Nvidia is actually suppporting previous gen cards with DLSS 4. Nice!