NVIDIA's Project DIGITS Will Change PCs Forever

[@CyberGizmo]

One thing I cut out of the video, is that I would assume Intel and AMD would be interested in this type of system, AMD currently holds the top spot for El Capitan on the TOPS500 list, and Intel has held that spot with Cray in the past with Aurora, I am intrigued by the NVIDIA DIGITS machine, but as an architect you should know, I never choose a solution until I have defined the problem, and those of you making comments about trust, and reliance on NVIDIA alone are correct, and I apologize for cutting the segment out how I would see this being an open architecture, not a closed one which only allows NVIDIAs machines as a participant.

[@andikunar7183]

With its RAM (not HBM but 8x LPDDR5 chips) it's more like a faster-GPU version (of a non-existing) Apple M4 Max Mac Mini, has a similar form-factor (hopefully also thermals) and is priced accordingly. It's not a shrunk A100 with 128GB and has only 1/4 of its RAM bandwidth and likely strong thermal limitations to also (

[@vincei4252]

If it requires cloud integration and numerous hardware and software licenses that can expire or make the hardware and/or software obsolete or unusable whenever the manufacturer says so then I'm not interested. That's how I reclaim my sovereignty.

[@ConorHanley]

Yep, thinking Nvidia is going to unleash anything we the people, or a bot near you, will benefit from without being screwed by Nvidia is living in a different dimension to me.

[@vincei4252]

@@ConorHanley I have purged my existence of every single subscription that was bleeding me dry. Even software I've used for decades like Mathematica that has now gone 100% subscription is at a dead end for me. I will keep using my perpetual licenses into the future even if that means running them in virtual environment.

[@ColbyWanShinobi]

💯

[@dieselphiend]

People will be working towards making it a more open platform. Even if it has to be hacked a bit to make it that way. I wouldn't worry.

[@r3dshed]

HaaS - Hardware as a Service

[@guilherme5094]

After the CEO of NVidia said that no one should waste their time learning programming anymore and that thanks to AI everyone is a programmer now, I really stopped paying any attention to this company. I just don't trust them anymore.

[@CyberGizmo]

I understand Jenson has a habit if saying lots of things which arent true...

[@muhdiversity7409]

@@CyberGizmo The more you spend the more you save. Must be nice living in a bubble.

[@alexxx4434]

You don't trust ANY corporation

[@chrisrogers1092]

@@guilherme5094 you don’t think this will *eventually* be true?

[@thesaltyone4400]

As someone who could never code the projects Cline (3.5 Sonnet) could code for me as fast and well documented would take me months of constant trail error and learning. Just because say you know how to code and develop and don't use AI isn't gonna stop someone like me from coding with AI or even using AI to make other AI So yes you don't have to be a developer I can make robust working applications with AI already

[@cyberhard]

It is Linux first because Nvidia knows anyone serious about AI does it on Linux.

[@stepannovotny4291]

The main acceptance friction point will be the complexity overhead of individually packaged compute modules. These need to be more like PCI cards in a chassis, rather than interconnectable devices each in their own case and power supply. Maybe NGFF would be a good platform for this? The hard drive model is a good one.

[@edemkumah5248]

I had exactly the same thoughts, when I saw NVIDIA's project DIGITS. Imagine 30 million people around the world with DIGITS connected. I wonder what amazing things people will design with that much computational power at almost no cost

[@HorrorAlgorithm0x1337]

I took a shot of whiskey every time someone mentioned AI during CES 2025... I was on my third liver before lunch!

[@jamerfunk]

lol

[@rhadiem]

Welcome to the present.

[@esra_erimez]

I thought the same thing about NVIDIA’s Project DIGITS, but the price tag....

[@CyberGizmo]

lol yeah but buy a RTX 5090 you are 2/3 the way to the price of that DIGITS

[@cacogenicist]

​@@esra_erimez - The price is not crazy, considering the hardware. Surprisingly. But maybe in a few years you'll be able to pick one up for $1,000.

[@POVgames]

@@esra_erimez I think the price is great!

[@nathanbanks2354]

The Mac Mini costs $2,649 and it only has 64GB of RAM. I expect the 5090 or 5080 to be faster, but they're limited to 32GB or 16GB. DIGITS has 128GB. A DIGITS is 1/20 the price of a GB200 which has 192GB of RAM.

[@NakamuraSatou]

Wait until nvlink 2.0 come out and your personal super computer is immediately obsolete

[@yuugenr7549]

Whoa then I'll wait.

[@williamlee2112]

I'm a supporter of AMD because of the open source nature of their ROCm stack. Howerver.. AMD should be scared shitless of Project Digits and the GB10. Project Digits could easily become a new super computing brick or AI module. AMD has a UEC NIC that does 400 GbE so they could compete with Nvidia's ConnectX with that. But they badly need an answer for the GB10 now.

[@XxSpYxX]

They have Strix Halo with up to 128GB MEM

[@andikunar7183]

I'm also really looking forward to this! For me, it's software-stack (CUDA, containers,...) really sets it apart from current offerings in this area (Apple Max/Ultra MacStudios). Its limitation is likely its memory-bandwidth - 8x LPDDR5X probably gives you 550GB/s, or 1/4 of an 5080 or A100. But this is a compromise they need to make, in order to give us 128GB at this size and price. It's really positioned price/performance wise against a very similarly capable Apple M3 Max 128GB MacMini or Studio. But with a much better software-stack, probably more capable GPU and NVLink. Another unclear probable limitation is the thermal envelope of its cooling. Apple spent A LOT of thought into cooling their Ultra MacStudios - and Project DIGITS hopefully has a more capable GPU than an M2 Ultra, even though it's RAM-bandwidth might be lower (512 instead of 1024 Bit wide). A future M4 Ultra Mac Studio might perform better (Thermals to use the GPU fully, higher memory-bandwidth, Thunderbolt 5 link), but Apple's software will be limiting it. The missing Apple GPU-virtualization is really bad, even Windows can virtualize CUDA nicely through WSL2 for Linux and container software. And with its Pro-chip MacMini price, it's very competitive to buying MacStudios with Max/Ultra chips.

[@AdamRogers]

When they opensource the software it will be great. Until them run AMD.

[@andikunar7183]

@@AdamRogers I don't understand. To my knowledge, only Apple has comparable hardware to Project Digits (500GB/s RAM + GPU). While Apple open-sources their MLX ML framework, it relies on their MPS underneath which is very likely to remain closed-source, and not able to run in a VM/container. Project DIGITS is the first step towards a solution with a better software-stack. I don't know about anything comparable from AMD.

[@Shreyash-t7m]

Hey I just saw something in you. I saw a old man ❤ but with a heart of a young boy who is so excited and fascinated about PC & Tech, And always will.. Love you brother.

[@michaelhuizinga1652]

Hey DJ I stumbled across this video as I was interested in some of the reveals at CES. Have you heard of ICP? The Internet Computer Protocol? It is a crypto network that is basically exactly what you were describing in the second half of your video. It is basically a cloud computing network that has distributed node machines across the planet that have some pretty beefy hardware specs. I believe if there is any hope of regaining individual sovereignty on the web it would have to somehow combine machines like Nvidias 'Digits' and a decentralized network like ICP. Curious what your thoughts are! Maybe make a video about it.

[@CyberGizmo]

Hey Michael, thanks for the great tip, I had not heard of it before, I thought at first you were talking about the Internet Communications Protocol, we are a sea of reused Acronyms is seems, and just looking briefly over it, it has some of the things I am looking for, so I will look deeper into it, and yeah maybe do a video on it.

[@chrisrogers1092]

Networks are decentralized by default. Especially the internet. I think the dark web (Tor network) is an already existing example minus the crypto part.

[@joseluisvazquez5126]

I share the idea that we are at peak centralisation nowadays, similar to the mainframe era. The cloud is the new mainframe, and the PC era that came afterwards could repeat again as another PC era, but this time "PC" means Personal Cloud. We do have the technology, that is not the problem, the challenge is to replicate the convenience of the cloud for normies without many tradeoffs. As nerds we do not need the same care, we can find our own way, but normies will just pick what is available up there and will prefer convenience time and time again. And their choice will shape the defaults and the industry. Imagine instead: - On every household to have the option that your internet routing device is a fully Open Source box. Normies do not care about it being open source other than having some guarantees that they actually own the thing and is working just for them, not a corporation that box is constantly calling home to. And they have that guarantee because us, the open source community nerds, have verified that the box does what it is supposed to do and its secure enough. - That box includes a Tor router facing the Internet, so the normies using it can chose to seamlessly have all of their traffic be anonymous by default, as a personal VPN, and also contribute to the Tor network for other users to route traffic through the external interface of their box. A Tor edge network. - That box can also be a NAS, have a HDD or SSD big enough to avoid them having to use any cloud services for storing their data backups and media. But still the UX is similar to using a 3rd party cloud retail service such as Google, Apple, etc. -- The normie phone can also be open source (eg Graphene OS) so that the user actually owns the device and is in control, rather than be "parentally controlled" by the phone vendor, as we have Android and iOS phones working today. The basic model is no information leaves the devices owned by the user, unless that user explicitly decided that it did. And the business model is good old hardware sales. From time to time, you might run out of disk space in the net box, or your phone will fall into a pool or get behind the times, or you want to have more redundancy, so you by an extra NAS only box that you install at home or maybe at some friends or family home for a disaster recovery scenario. Not subscription fees, no tracking, just reliable devices that sell like hot bread.

[@nathanbanks2354]

It looked like an overpowered Jetson or underpowered A6000 to me. Like the Jetsons, it has enough RAM to run larger LLM's or diffusion models, but I doubt it has enough power to do it quickly. However it would be much faster than an m4 Mac Mini with 64GB of RAM, and the Mac Mini is $2,649 with half that much RAM so the price is similar.

[@rustyroy5385]

I'd imagine it won't be as powerful as a 5090 but has the benefit of being able to run larger models due to the larger RAM capacity. No doubt NVidia would like you to get both - 5090 for running smaller models quickly and Digits for running larger models a bit slower.

[@henrytuttle]

I bought my first Commodore 64 in the winter of 1982. It cost me $600. That's supposedly around $2000 in today's dollars, but I think you can buy far less now than back then for $2000. Anyway, that's pretty close to the $3000 that the Project DIGITS machine would cost. I think buying a Project DIGITS machine is pretty close to the equivalent of that C64 in where it stands in relation to the market and where we'll be going.

[@jessehorstman]

Why are you playing 5 minutes of nothing at the beginning of your video? You have something interesting to say. Thank you. Please say it.

[@Silent1Majority]

How many Nvidia licenses will I need to purchase along with this mini-workstation ? This is my concern.

[@yuugenr7549]

I needed this since last year. I hope it works offline. I dont want anyone stealing my projects that im going to use it for. If its offline than it can work for me safely.

[@jessehorstman]

bus speed and the limitations of latency and parallelization have been limiting the applications of computer clusters. It's a good goal and we have seen steady advancements toward it. The unfortunate situation regarding older components is that they become relatively impotent and rather than contributing to output in a distributed system they tend to reduce efficiency. On a shorter time scale, having the ability to easily assemble a performant cluster of computers and scale it to meet imminent demand would be great. Hopefully Nvidea isn't the only option.

[@freehaven-junprince2376]

My god how are more people not talking about this? I was planning to get a 5090… I think I might pass on it and wait a few more months

[@DJNicke]

Using quantum networking (quantum entangled photons carry traditional encryption keys that are invalidated if looked at by anyone other than the intended recipient) on a platform like IOTA's tangle for decentralized and transparent data transactions.

[@terrydaktyllus1320]

NVIDIA products and services are boycotted from my existence until they take Linux and Open Source seriously. I don't play nonsense post-2010 AAA "games as a service, gamers as whales" games anyway, everything pre-2010 I do play works fine on Linux with Intel and AMD GPUs that do take Linux seriously. I also have no interest in AI, especially NVIDIA's AI, because I am not ready to turn my brain off and be a corporate drone just yet.

[@johnkost2514]

You can cluster more than two of these by plugging them into a Nvidia ConnectX switch (100-800G ports) for a nice onprem AI inference (or training) facility.

[@henrytuttle]

The crazy thing is that it does 1 PFLOP. The #500 computer on the TOP 500 does 1.022 PFLOP! That would make this the #501 fastest computer in the world! I suspect that the TOP500 list has some additional considerations or is leaving out a lot of computers. Still, it shows how incredibly powerful this tiny, relatively inexpensive machine is.

[@rhadiem]

Guy, this isn't designed to "GROW WITH US" It's an all-in-one. No user-replaceable anything. (ok maybe hard-drive..) You can do all that this does already right now with a good PC or Ebay GPU server. Yes this is smaller, cheaper, faster, but it's not new to those who have been in the homelab or commercial space. It's just cheaper for peasant end-users to pay attention to.

[@carpetbomberz]

Scalability reminds me a bit of Beowulf clusters. Just keep adding nodes. Sounds good to me.

[@taylor-worthington]

Interesting video. Interesting thing. A rack of 4 for $12,000, but might come down (or be scalped for 3x, like GPUs). Wouldn't a "real" device be more concerned about cooling than its looks? All of a sudden, a new device means the technology is so incredibly different that Coreboot is out of the question. (Look mah, still just called a computer.) I think the hope for FLOSS people would be if it can run any linux distro and stack, but seriously it's very meaningful that it would include coreboot/libreboot/etc. What good is the rest if you don't have booting? "Increasing commitment to open source" means more lines of code are taken for free while they tighten grip over booting , online dependence, or licensing on something else. Can it run a full FLOSS stack? may it? will it? will it for long? Is it future proof because it is the butt of Moore's law? I don't get it. Why is it future proof? Modularity is seriously awesome but as it did 25 and 50 years ago, the "bus speeds" will be a tradeoff compared to putting the chips right next to each other.

[@PeterKese]

It is an AI supercomputer in terms of compute performance, but it has the same memory-bandwidth as MacMini (8x LPDDR5), or slightly less than GeForce RTX 4070.

[@mattelder1971]

They aren't changing anything with it until they at LEAST cut the price in half.

[@keratishvili]

Great Content, not many people speak about it

[@DLLDevStudio]

Running a 400B model requires around two DIGITS (or so they say). Even if it needs double that, it’s still super affordable! I hope that’s true. Currently, you need to invest at least $50k to $100k USD, and that’s if you’re willing to go for used parts, to get a 400B model running at a bad tokens-per-second rate, especially with proper quantization. That’s horrible! If I could buy four of them for about $12k USD each to get it working, I’d do it instantly!

[@Alphadec]

This sounds scary. Prefer to have hardware that I control.

[@opius1199]

I see a rise in thinking where we can use existing hardware to create our own domain again, where big tech is not breathing down our neck and putting their sweaty hands all over us. Which is a good thing. Unfortunately the biggest issue we are facing with decoupling from the existing, is where all the good intentions are turned around to be used by bad actors. If we are going to succeed, there needs to be an overarching philosophy in creating this new space where freedom comes with responsibility. Meaning a new protocol for communication also needs to come with identification, linked on the hardware level and with good enough encryption to not be bypassed. Your outline of a new future is all good, but unfortunately I see us going down that route and ending up with a lot of tools that all the criminals want to use. Like we have done in the past. When burden of proof is on the "good side" in most legal systems, that also means prosecutors/law enforcement, etc... ends up spending 100 times the amount of time on getting even the most clear cut cases prosecuted, compared to the defense. In the technology domain, we need to start focusing on this aspect as we go into a new future. Privacy is a luxury on todays Web, we have to make it so privacy is exchanged with accountability. If big tech is gone from mining our data, you should also be able to be accountable for what you do on the Web.

[@donnydarko7624]

Pretty sure you're not going to not going to see any windows support for it ever. It's going to run it's linux fork called DGX OS. That way they don't have to deal with any of the bloat that causes so much inefficiency on the windows platform.

[@Enraiful]

>Start DJWare video >Blasts out of my speakers Holy based, getting tired of quiet youtube videos.

[@AdmV0rl0n]

I always enjoy your video's and comment DJ! I think that we are in a battle for the soul of our compute and data. But we are losing and losing very very badly. I have a Jetson NANO device. I've seen this playbook before, and its not pretty. It runs 'their linux' and it has some hardware tie ins to 'their hardware', and its a total linux dead end at Ubuntu 18 because they dropped it like a stone. That whole device looks to be very dependant on an Nvidia software stack, and I would be surprised if chunks of it don't need to pay license fee on usage. Rather than getting things back in our hands, all I see is amazing compute - yes - but its very much tied to the big player. Nvidia have that fat share price right now because their compute is epicentre in AI. This is an embrace / extent out of data centre, into edge or 'home', but it has baggage. If Nvidia wanted you and I to have access to AI compute, they could put decent quantity of RAM on consumer GPUs. They don't. And that is deliberate. Here is what I think the state of things are at though. I think this wave that we are bang in the middle of, is not defined. I think no one really knows how its going to go, including Nvidia, and that mooting that - never hand a gun to people unless you know who its going to be pointed at is a good call!

[@Iswimandrun]

Federated Distributed Secure Compute "FDSC"?

[@jjhw2941]

They are using Linux because they already have the entire compute stack on their Jetson line with Jetpack, there is supposed to be a Jetson refresh this year with the Jetson Thor which may or may not use the same chip.

[@oscarcharliezulu]

Secure distributed ledger like block chain and without the fear of compromise? lol that’s not the block chain I see!

[@TheLinuxGallery-qz2vs]

The only thing I saw that makes any contribution towards the utility of the computer is the 20 core CPU. The thing that makes GPU's relevant AT ALL is the fact that if you try to extend your processing power with more CPU's, they will bottleneck each other and result in a net loss of performance. An extremely large consumer-grade processor like that could eliminate the need for a GPU. Every other issue - including those solved by the net-fs concept that makes the DIGITS machine so extensible - seems to be, to paraphrase Rene Rebe, a problem of priorities. Devs of "THE" linux projects COULD fix every problem they have, like Google and Netflix did for themselves using their code. They could make thinkpads extensible in exactly the same way as this - either through Google's net-fs idea, or some form of RAID solution, or something completely different. They could do that in the early 2010's. As long as they just don't wanna: even when people do it apart from them - it doesn't get done. The first thing that comes to mind is how KDE deliberately makes their desktop trash can hard to use - but the plugin system wants to use kernel-space for no particular advantage, which I only found out because an un-modified version can't run anything that uses plugins on freeBSD. Their project, but not their problem. They're going to just do what they feel like doing and ignore what they don't care about - even though they're the most well known FOSS desktop in the world, with the best position to shape the industry. (who, by the way, worked extremely hard to position themselves as the first choice of the general public)

[@modolief]

12:34 - some practical examples

[@camronrubin8599]

Youll be able to double the performance for half the price when it hits ebay

[@radeksparowski7174]

reminds me of the HPs project "the machine"

[@radeksparowski7174]

and combined with the good old beowulf cluster, where some lab dude chained together idling lab workstations with one in charge the others shared ressources to get one supercomputer

[@ericneo2]

Lol that Win95 cd sampler ending

[@boardsontt1756]

A lot of “doomer” bots in the comments. Either that or a bunch of butt hurt, would be devs 😂 I love it!

[@ConorHanley]

Trusting corporations to serve us..what fool ever thought that?

[@_neon_light_]

I look forward to 6 to 8 years from now when they start to show on ebay for a fraction of the original price. 🤓

[@peterwassmuth4014]

Awesome! Thank you for Sharing! 💯✴

[@leaveempty5320]

IT goes round another cycle.......

[@terrydaktyllus1320]

So name one of the other cycles you are talking about because I don't get the point you're making - and that's someone who has worked as a techie in IT for probably as long at DJ has.

[@CarlosValero]

I want two of these machines!

[@Bluesourboy]

I think people moss the fact that this might mean better integration with Nvidia and Linux.

[@kevinsteinman8967]

Perhaps Nvidia and your wallet is more to there calling. It's strange that Nvidia has all this interest in Linux all of a sudden.

[@tonioyendis4464]

Very kool Santa Claus beard!

[@williambrasky3891]

What worries me about this is the power, the required power. How much does one of these boxes pull from the wall? Is it really scalable and democratized if by the third or fourth box you add, you have to wire up a dedicated 220V circuit?

[@GeorgeSomething5]

Another great video about the future and potential, breaking free from the homogenized system we have found ourselves.

[@tibbydudeza]

The Apple Vision Pro of workstations - rather would invest a Mac Studio Max.

[@janmagrot]

Yeah, whole jetson family runs on linux. There is whole comunity, that plays with this on linux for years, already.

[@camsand6109]

Exactly wheat ive been craving. I hope you’re correct

[@shrek22]

8gb ram is not much. I’m sure there will be a way to add more.

[@bgone5520]

so i havent bothered to watch the video but are nvidia charging so much for GPU's now they will accept fingers for payment

[@rickharold7884]

Looks super juicy. Love it

[@CyborgZeta]

I personally find it difficult to trust NVIDIA. I do not buy their GPUs.

[@terrydaktyllus1320]

Neither do I - overpriced and proprietary rubbish to play derivative and buggy "games as a service".

[@shrek22]

Look at their market share though. In the datacenter, they engineered their way to what we know as machine learning. Something amd didnt do. I’m biased towards amd as well. But invidia has done incredible things it seems.

[@ණChỉYêuMìnhEm]

So?

[@dreamphoenix]

Thank you.

[@ash1kh]

Nvidia? Nah thank you. I am good.

[@elalemanpaisa]

Uffff.... Mister Ware, I have to disagree for the very first time since you got white hair. We haven't been doing it differently for the last 4 decades how this systems appears to me.. Need more power? get a new server and hook him up with a fast connection. If this scales it has to be concurrent workload - meaning you can slice the problem in smaller packages and distribute them. this is nothing new. If I look at my closet super computer I can hook more over TB4 it it is even overkill to have efficient workload distribution I'd be fine with Fast Ethernet (that is 100M)

[@zcavaleiro]

Wise "words"!

[@sleeepykat]

MTPA!

[@alexxx4434]

DIGITS made be a corporation with proprietary technology will set us free? Um, somehow I'm doubtful of this premise. Also, PC became a widespread thing only because IBM have lost control of it.

[@fontenbleau]

Finally.

[@pilotbsinthesky3443]

Just imagine if all these Nvidia haters were in charge of tech when windows came out. their boss asks them why they didn’t implement windows at their job and they said “because I hate Microsoft” guess what you’re fired! You can either embrace the future or get left behind. Sad to see emotions get in the way of using your brain.

[@rodfer5406]

Wow

[@ණChỉYêuMìnhEm]

Now I can build and scale my data center aha