What is an NPU and when will it matter?

[@Ucfahmad]

I don’t think viewers are against ads - we understand that it’s a necessary reality. We just want transparency - we want to know if what we’re watching is an ad, or organic content. When you blur the lines between content, sponsored segment, and product showcase, it honestly all just feels like an Ad. And in this particular video, the history and NPU explainer sections feel like they were built around the ad, rather than the other way around. That’s my (hopefully) constructive criticism.

[@maxpowers2168]

I stopped watching when he said "Today's Sponsor"

[@BrianUnderwoodCodes]

I think that the big problem for me is that I come to Snazzy Labs because I often see more advanced geeky things that I don't know. There's some trust there, and so if I see a video about NPUs I expect that I'll learn something new or interesting. But I didn't really learn anything new, and so that makes the whole thing being an ad feel that much worse...

[@Sevenfeet0]

I agree. It's not like you haven't had plenty of sponsors before. It's just that this one was a weird "uncanny valley" of something you would not expect in a Snazzy Labs video. I cocked my head in confusion when the CPU graphic explaining all the parts then went to a SnapDragon logo over the whole thing. I'm thinking, "why would a usually Apple-centric channel decide to use that graphic when.....ohhhhhhhhhh!" I've seen worse.....a lot worse over at some sponsored ads at MaxTech.

[@BrianUnderwoodCodes]

Kudos on the most recent video, though. That’s the sort of thing I’ve been looking for for a long time

[@haydensimpson02]

From my point of view, the video topic was not prompted by a specific company and the advertisement is for a product related to the subject, but the subject is not on "the best AI hardware" but rather "why AI hardware exists". In other words, the video was not an advertisement like people are stating, rather it has an advertisement in it. The video topic was not prompted by the sponsor.

[@markjstradling]

What would demystify stuff is saying “you’re going to waste 10 minutes of your life not getting an answer to the thumbnail while I sell you something”

[@JamesR624]

Yeah, I used to love this channel but been noticing more and more of.... this, lately.

[@mikkelens]

A lot of tech youtube is waking up to the fact that most of the content here reads a LOT like an ad. If not directly paid for, thats still a low bar. We want critical journalism and not promotional vectors.

[@clementcollier8432]

Like most of the youtubers he overextended and now has to pay for a lavish lifestyle AND long-term commitment for the "business", he has salaries to pay and he just bought a house! I do think it's really bad for the consumer/viewer. The benefits of KZbinrs previously were that they were not indebted to anyone and thus could be free to direct their attention to whatever pleased them and largely say whatever they wanted without fear of repercussion. Being in such a need of financial results defeats the whole purpose of KZbin and they just become another extension of mainstream media, enslaved to the need of making ever more money. This is not desirable nor sustainable. You can't put much trust in KZbinrs after they have become "too" successful. He should just have kept the operation small and nimble, but then it would require more work on his part and humans really don't like that. The whole point is to have someone eventually work for you while pocketing the larger share of the profits, so the results are highly predictable. As for the subject matter, maybe NPUs will become more useful down the line, but then again it is very hypothetical, the whole thing was driven by marketing instead of real need, unlike the GPUs that were created to solve graphics for video games in the 90s. I don't think there is much point to them because GPUs can do the same work while allowing more use cases. They may consume more, but power efficiency is a marketing gimmick from Apple because it serves them well, in the real world most people don't care that much because it doesn't have as many benefits as the advocate likes to pretend. Having modular and generalizable architecture is more useful I think, especially has we been going against bad yield from large die areas, it's a tradeoff of dedicating so much surface to an overspecialized application.

[@Äpple-pie-5k]

tbh he's trying but didn't do his research and has no techie connections, won't even reach out to me to help him ... think of YT channels kinda like legacy media and entertainers in that they aren't going to do 3 months of research to give u the best journalism

[@Äpple-pie-5k]

@@mikkelens yep, truth and more quality content gets more viewers way more than sucking up the viewers u got for money-now, don't mortgage your future

[@hyenatron]

The irony of the section on accessibility having broken captions.

[@cinebenjamin]

So sorry, I fixed that portion.

[@YouArentValid]

What does it mean that a “portion” of the video is sponsored by Qualcomm? It seems pretty black and white, is it sponsored by them or not?

[@rikwisselink-bijker]

Some brands want to have the option to PR spin and say that they only sponsor the ad read (in case something in the rest of the video blows up). So this just means that the video was sponsored by Qualcomm.

[@fallende]

This channel’s gotten weird since a few months ago. Ad after ad, and much lower quality content :( I’ll give it an opportunity though because quinn’s great

[@snazzy]

2 out of the last 10 videos have had an ad…

[@fallende]

@ you are 100% correct. I’ve felt like after the “ad” about the robot vaccum cleaners the channel’s felt a bit off to me. Up until then, banger after banger. Idk what (if anything) you’ve changed in terms of the videos since then. Maybe it’s just me and if that’s the case I’m sorry :) Ps. I really don’t want to use internet anonimity to bash other people’s very hard work, so please take it as it is, (poorly written but) constructive criticism

[@iker42]

In the hopes that it helps things get better; I’ve considered unsubscribing lately due to the feel changing but Quinn is great so I’ve stayed for now. Idk what it is but something has changed.

[@betweenthepanels9145]

@@snazzyyour robo taxi video was a sponsored hype piece and this one is too

[@Äpple-pie-5k]

@@snazzy never give the feel you're trying to make money, it makes you less, get more traffic and let that handle it will only micro-nudging subtle factors for revenue

[@harrydrewer905]

I think you might've missed with this one. Next time be more upfront about the video being sponsored.

[@alexlock3176]

Yeah, no... this views straight up like a Qualcomm ad, regardless of the timeline of how they approached you.

[@MikeJackman]

is this whole video a friggen ad?! for stupid windows ai slop... good god... the mighty have fallen

[@mikkelens]

yeah idk how this fell through. I subscribed for more critical reviews than this advert. Not gonna stick around since I haven’t been impressed with the last couple videos in a similar vein

[@mikkelens]

It should be a made LOT more obvious that this video is sponsored by qualcomm.

[@DefendIntelligence]

Omg that's me on the screen :D thank for the shoutout :

[@atayklmaz7116]

This video is an ad

[@cinebenjamin]

Only a portion is sponsored. We tried hard to talk about NPUs generally, as many devices now have NPUs, such as MacBooks, iPhones, Android phones etc.

[@D3R3bel]

​​​@@cinebenjaminI think this video would have been way better titled as what do/can NPUs do for you, rather than "why we need them". The whole video feels like an ad for NPU copilot pcs because it's trying to tell people they need something when all the feature sets showcased here will be at best of extremely minor interest to the average consumer. Frankly, it kind of reads like a sponsored blog.

[@cinebenjamin]

​@@D3R3bel We've had this video planned for awhile even before Qualcomm was attached. NPUs are found in both Windows and Macs, so we mostly used examples that were cross platform, Cephable, Davinci Resolve, and LMstudio are all cross platform and apply to all NPUs. The ultimate point we tried to get across was that while we have NPUs the software isn't ready to support them yet.

[@Im_Behind_You]

@@cinebenjamin No, you don't understand. It's an ad because @atayklmaz7116 said so

[@spht9ng]

anti-AI obsessed youtube reactionaries try not to write off innovative tech challenge (IMPOSSIBLE)

[@chetanvmalhotra]

I was impressed by the video. I learnt new things. I came to the comment section and saw people were angry with this coverage. Great video Quinn!

[@FoxxFire]

When you said Qualcomm was the sponsor, I literally thought you were joking for the first few seconds. I have to say, I love how you took a sponsorship opportunity (which you of all people deserve; your videos are top notch and you should make as much money as possible to stay on KZbin for all our benefit) and turned it into an educational video.

[@emeraldcityevan]

Starts off interesting, turns into an extended infomercial.

[@liranlatin1]

Great ad… I mean video,

[@Telzrob]

Next time lead with the sponsored video notification. It would have saved me seven minutes of my life.

[@GuislainBonnard]

The beard is magically back?

[@b0ss__li]

AI generated I assume

[@chidorirasenganz]

@@b0ss__lithe beard is a hallucination

[@cinebenjamin]

Was just filmed earlier than the last video. 😅

[@chidorirasenganz]

@@cinebenjamin the boring realistic answer

[@dmug]

The answer should have been the NPU was used to object track a 3D beard and attach to the video.

[@benjaminlynch9958]

I honestly think NPU’s are a fad that are going to disappear from computers in a few years. They’re GREAT for mobile devices where every watt matters, but for computers, discrete graphics cards (both desktop and mobile variants) offer orders of magnitude more TOPs performance than the NPU’s coming to market. The Qualcomm NPU in this video can do 45 TOPs; meanwhile the new nVidia GPU’s that came out this month can do 1,800-3,300 TOPs. There’s zero chance desktop PC makers are going to waste money on an NPU when they’re already delivering orders of magnitude more performance. And without desktop support and only putting them in laptops (where the value proposition is better but still not necessarily great), software vendors will continue to prioritize GPU support over NPU support because GPU’s - even integrated graphics - are in every device.

[@monstercameron]

you committed the cardinal sin of NPU perf numbers, 45 TOPS at what datatype!!!!! Where is Ian Cutress when you need him.

[@jcfawerd]

He is busy defending LTT

[@Jonteponte71]

Meanwhile, people whine about Apple being "behind" on AI when they introduced the first NPU on Apple Silicon in 2017.

[@middle_pickup]

Lmao he did an ad for co-pilot PC's! Unsubbbbbbed

[@blabberingblab]

Silly Snazzy, this is GPU day!

[@hookey1000]

Snazzy Ads

[@yoram_snir]

Sales pitch for dragon 👎

[@xmaverickhunterkx]

It is funny seeing this after DeepSeek.

[@TheCerealHobbyist]

This felt like an internal corporate hype video or an Apple event.

[@Äpple-pie-5k]

i thumbs-downed this video and i don't hate u but something is off here, i'm not a clown viewer but could teach you on your own topic... please squirrel home on your methods when u go out of your field of comfort and, extrovert as u are, reach out so it doesn't seem like someone who did 3 hours of research on what u don't know about to then give some instructional video to everyone else because u have 3 more hours of experience research, wtf

[@Snowsea-gs4wu]

At 2:20 do a collab with Adrian's Digital Basement. He is a wizard repairing Commodore 64s and both will gain subscribers! Thanks for the video!

[@garethsnaim8174]

Jesus suffering should have read the comments.

[@tormaid42]

Not the disability washing of AI…

[@zwparchman]

"... probably in your hand" ... Sir, this is a bowl

[@JamesR624]

"Computerized Telephone" is my new favorite old-timey word for smartphone.

[@samuelbock8550]

New Snazzy makes my day snazzy :)

[@bluekeybo]

1 Commodore 64 was hurt in the making of this video

[@Hbomb117]

Sorry Quinn but this vid sucks man

[@rogerhuston8287]

Are NPUs better than GPUs? All I hear about are NVIDIA GPUs and AI.

[@ionaspolitis]

Loved the intro 😂

[@michaelbabiuk8758]

How would you classify Apple's R1 chip in its Apple Visio Pro headset? Is the R1 chip essentially an "external" NPU working with the M2 SOC in this headset?

[@haydensimpson02]

You probably wont see this, but idk what people are on in this comments section. I was headed down here to say how informative this video was! Before this video, i never saw any demonstrations of just how useful it could be to have an APU. I don't think the sponsoring was out of line at all. In fact, it doesnt seem like it changed any aspect of the video except for what hardware the demos were on, and the focus of the video was certainly not on hardware. Thanks for another great video!

[@SwiftyRob]

Quinn has unique videos and I only see big fans commenting out of frustration, how about given he’s soon to be a father we see this video for what it is and all support him, so that he can continue to bring more unique videos one day into this generic tech KZbin bubble.

[@AlanTheBeast100]

The C64 had a fairly advanced graphics chip for its day. Sprites anyone?

[@Vertig0ne]

The truth is, a CPU is for sequential tasks and a GPU is for tasks that can be parallelised... So what exactly is an NPU used for because you seem to have confused NPU for a GPU

[@dctaken]

Will there be a video on fixing that C64?

[@OldMan_PJ]

KZbin already auto-captions and auto-translates all videos, I watch Finnish and Japanese content regularly with English subtitles. The channel doesn't have to do anything, it even works on child accounts. You forgot to mention the subscription fee companies are going to charge for their AI "features" that run on your own hardware.

[@akr4s1a]

I think its a miss to not mention that some processing still happens on the CPU and GPU, there is a lot of offloading to the NPU but not talking about the Adreno usage increasing albeit mildly in some of those workloads was a miss imo

[@wileysneak]

what so, basically no programs use the NPU and when they do its slower than CPU? qualcomm can keep their beta hardware

[@cinebenjamin]

Sort of, but they are faster when using the NPU for things that can excell on the NPU. Like davincis magic mask, is several times faster than running on the internal GPU.

[@daviyen]

“Just a few watts of power” … I’m sure that’s why it’s plugged in right now

[@willfancher9775]

Don't discrete GPUs basically always have a pretty beefy "NPU" on them as well? Like nvidia's tensor cores I'm pretty sure vastly outperform these new integrated NPUs, and do it pretty efficiently too.

[@cinebenjamin]

NPUs have an extremely stripped down instruction set. It saves energy and space on the SOC.

[@Im_Behind_You]

Love this video!! But it just might be because I'm not knowledgeable enough to nitpick all your mistakes and inaccuracies

[@LukanSpellweaver]

gross, an ad about ai

[@lenn55]

Your voice sounds different in the parts of the video where you're sitting behind the desk.

[@Äpple-pie-5k]

It matters already and Apple believe it or not was first, just flopped like crazy in taking first mover advantage.

[@Slimothy]

*It's spelled "NPUs" if you want plural. Your title and thumbnail with "NPU's" doesn't make sense. Just FYI.*

[@rynabuns]

Good heavens! Sipping tea from a glass in lieu of fine porcelain-unthinkable!

[@Dr.Dawson]

You actually don’t want to know what was in my hand while watching hehehehe.

[@owlmostdead9492]

I don't need an NPU, maybe in a couple of years when Ai is something more than a gimmick/meme generator

[@chidorirasenganz]

@owlmostdead9492 already is

[@spht9ng]

It already is, you're just living under a rock.

[@owlmostdead9492]

@@spht9ng I'm sorry but I will assume you don't even know how an LLM works, I do and run my own LLMs locally, on my homelab cluster. They're overhyped to the moon, there is no intelligence, no thinking, no "emergent" behavior and never will from an LLM.

[@tigersusyt]

My phone with an SoC from 2018 has an npu

[@tipoomaster]

Bu-Ton

[@mickcolips]

Another Buckaroo in Town… Can’t keep pace with U. Its getting Silly Con My Commodore sends U his thanks for the memory.

[@RobertOortwijn]

Did you get a new microphone Snazzy? Your voice sounds different from what I'm used to.

[@trevorbeingtrevor]

Yes something was off to me about his voice and maybe other parts of the video as well at points. Especially some of the voiceover over some of the b-roll. 🤔

[@googleyoutubechannel8554]

is 'npu' still a thing, have we figured out that it isn't different enough from a gpu to need a new marketing buzzword?

[@kenjain]

One of your best cold opens so far

[@mirzaangon]

The beard is back…?

[@2brave1]

It grew back pretty quick! Probably because he’s so virile.

[@emmabentley7945]

Yay ! the beards back..... Next week , noooo it's gone 😂 , fantastic video as always, Very informative ❤

[@LH782]

Real-time translation subtitles are horrible. I'd much prefer the system looking ahead and giving, say, two lines of text at a time instead words jumping around one-by-one.

[@DanielSanNinja]

thanks youtube for ruining it for everyone... autotranslated title .. aaaand always some weird audio language ai translation... yeah nobody asked for that and it's annyoing af

[@tanmaywho]

Ok maybe I ma speed, this being the second video I have clicked withing sub second.

[@madw2615]

this may sound like a dumb question but can an NPU be used for any other useful tasks that is not AI / machine learning if not for me IMHO it is going to be like a 3D movie pointless

[@chriswatt2702]

Probably the best task demonstrated was in video editing. I am guessing photo editing will be the second task that could take advantage of this feature.

[@eazyliving]

Well, there should be a C64 repair video with Adrian Black...

[@pumpuppthevolume]

useless extra chip .....am I right

[@vhol93]

Lol that intro

[@GEN9100]

This is pathetic misleading commercial for snap dragon, an rtx 4090 hits around 1.3k tops. And tops doesn’t show the whole picture. How the hell can you say gpus cant handle it. When in actuality they are more capable than a stand alone npu.

[@bviptech]

This is massive fopr someone like me who is vision impaired. Any pc i find the buil tin accessibility rubbish but the third party software is use destroys the pc memory and graphics. Its nowhere near as bad on mac as the built in stuff is pretty robust and can handle a majority of my daily tasks.

[@jaw2112]

English accent 2/10

[@JimmyLamStudio]

that monocle is 3D Printed, EXPOSED

[@sledgehammered1765]

but the scotch on the carpet

[@AnonYmous-yz9zq]

Its way cheaper to stay a couple of years behind the leader,s call me after the apps arrive. About two years after.

[@chidorirasenganz]

You should try out Private LLM instead of LM Studio

[@atakolday9066]

YOU CAN’T JUST SAY PERCHANCE

[@TheCrunchLog]

Genuinely a pretty cool video, I had no idea some of the use cases of NPUs. Pretty neat! I can see why they’re here to stay for the foreseeable future.

[@robbyj4wmee]

Becoming more of an infomercial.

[@Ve55el]

...ADR or new mic?

[@Ve55el]

...someone forgot to hit record huh?

[@Ve55el]

....maybe raising the high end eq will give that "room" feel.

[@cinebenjamin]

Quinn was coming off of a cold at the time of recording, so that affected his voice.

[@username9774]

now with deepseek everything else seems useless

[@chidorirasenganz]

@username9774 not really R1 is almost on par with O1 which is beaten by O3 and Deepseek + local AI setups still lack UX features like canvas, voice, image recognition and project support

[@eddietours3728]

hi what speakers are at the back ?

[@googleevil]

I want to see you in 4K 60FPS HDR.

[@MartinNelson]

Right off, I disagree - I don't think Machine Learning is here to stay. In fact, given the large-scale outcry from creatives and small businesses over the new copyright and intellectual property loopholes that Machine Learning and AI tools utilise - I genuinely think we will start to see a rejection of these tools. I'm seeing lots of people around me FINALLY move to LibreOffice and away from any software package that includes AI for these reasons. Search around and there are endless threads on people asking how to remove Copilot, there are people asking how to know if a book has used AI to be written. Given the widespread copyright implications of any AI assistance on operating systems now, I'm even seeing non-techy friends and family asking how easy to use Linux is. We're at a tipping point, and I don't think that AI and machine learning will win for large scale mainstream use. Don't misunderstand, I think that machine learning has great potential in a number of fields, but for mainstream usage - I just don't think it is the certainty that many companies anticipate. The EU is getting a lot of fire over their AI policies. Keir Starmer in the UK is getting ridiculed for his stance on AI being able to make Britain a global leader. Deepseek has effectively just wiped out any reason OpenAI has to exists (potentially because they used OpenAI to train Deepseek - oh the intellectual property ironies). If as you claim, you wrote this video before the sponsorship offer, I have to wonder how encased in the Tech world bubble you are. Outside of the tech bubble things aren't looking quite so favourable to AI and machine learning. I have lately been likening it to 4k displays or 3D (stereoscopic content) - a tech development pushed onto the consumer rather than a solution to consumer problems. And AI will go one of these two ways - like 4K it'll be a really slow uptake where people grumble over its usage and necessity (look how slow 4K resolution has taken to penetrate the market and it still isn't dominant yet). Or, it'll go the way of 3D (Stereoscopic) films - it'll die off as the fad and shiny nature of its presentation gives way to the flaws of the underlying technology. From what you've published in content to date, I honestly don't think you're naive enough to really believe that AI and Machine Learning are a mainstream certainty. If you are - then I'd turn some attention to the massive levels of anti-AI discussion occurring in creative spaces. Publishers and booksellers, clothes designers, movie makers, voice artists, podcasts, and loads of other sectors are actively and proudly marketing that they are not using AI. Meaning there is a sizeable market for products that have rejected these technologies. So it might be time to step out of the tech bubble and reassess if you've become stuck in an echo-chamber with these opinions.

[@melgross]

Maybe this video should be redone without the sponsorship, to be neutral.

[@michaelerfurt]

Can you please disable autotranslate. Thanks

[@EnlightenedSavage]

Whole lotta of fluff and very little sometimes erroneous information. Move on.

[@fergusyoung6782]

3:42 I feel called out…

[@NithinJune]

everyone in the comments is complaining but this video is like fine?

[@ps3301]

Npu is too weak to run llm though. They have to triple its size

[@axa993]

Eh. You could do without videos like these.

[@TerrySitu]

The first smartphone System-on-Chip (SoC) to feature a Neural Processing Unit (NPU) was Huawei’s HiSilicon Kirin 970, announced in September 2017. Details on the First NPU in a Smartphone SoC • SoC: HiSilicon Kirin 970 • Company: Huawei (HiSilicon) • Launch Year: 2017 • First Phone: Huawei Mate 10 series • NPU Functionality: Designed for AI tasks such as image recognition, object detection, and power-efficient machine learning acceleration. While earlier chips (like Apple’s A11 Bionic, also in 2017) introduced AI-focused components (like the Neural Engine), the Kirin 970 was the first to explicitly include a dedicated NPU for on-device AI processing. From ChatGPT 4o.

[@manojos4114]

Dwight?? is that you?

[@joeykeilholz925]

Never

[@natandse1287]

Its Natural Language Processing, not Neural Language Processing, and I get that you tried to make this video responsibly but this whole video feels really irresponsible. I hope qualcomm paid you well for this because this advertisement will make them billions.

[@vhateg]

NPUs are also on Intel, AMD and Apple CPUs :-) I genuinely feel this video is informative and decent, as I really didn't know what an NPU does. I don't buy into the whole AI thing, but I don't see how this specific ad would be in any way irresponsible. Qualcomm paid a creator to explain how NPUs work, lol.

[@natandse1287]

@ npus are also built to store your data in a way that’s accessible by any company that wants to buy it ignoring that in a video like this is irresponsible

[@cinebenjamin]

@@vhateg Thank you! Qualcomm only paid us to mention the X Elite Soc though. We spent a lot of time researching NPUs cause they felt like a black box to us.

[@cinebenjamin]

@@natandse1287 That is totally not true. You're conflating NPUs, and LLMs.

[@natandse1287]

where do you think the companies operating those LLMs buy that data from

[@NickWhitehead24]

NPU for me has always stood for 'Network Processing Unit'

[@alain-m2t]

AI NPU = surveillance = OCR your screens and other malicious stuff

[@Marauder-q2v]

You know you can also use your gpu for the exact same thing right?

[@alain-m2t]

@@Marauder-q2v Well, the recall system has more requirement that only the new npu would provide - why do you think all the push by nvidia, .. for new ai hardware?

[@Marauder-q2v]

@ power efficiency. NPU’s are more power efficient than GPU’s. But the GPU is more than capable of doing the NPU’s tasks. In fact Alchemist+ is way more powerful than any NPU in raw TOPs. Nvidia pushes AI because raster performance gains over each generation has completely leveled off.

[@JohnWick4K]

what a useless video with false advertising