Nvidia's Project Digits is a 'personal AI supercomputer'

I'm looking at my Jetson Nano in the corner which is fulfilling its post-retirement role as a paper weight because Nvidia abandoned it in 4 years.

Nvidia Jetson Nano, A SBC for "AI" debuted with already aging custom Ubuntu 18.04 and when 18.04 went EOL, Nvidia abandoned it completely without any further updates to its proprietary jet-pack or drivers and without them all of Machine Learning stack like CUDA, Pytorch etc. became useless.

I'll never buy a SBC from Nvidia unless all the SW support is up-streamed to Linux kernel.

I feel this is bigger than the 5x series GPUs. Given the craze around AI/LLMs, this can also potentially eat into Apple’s slice of the enthusiast AI dev segment once the M4 Max/Ultra Mac minis are released. I sure wished I held some Nvidia stocks, they seem to be doing everything right in the last few years!

Nvidia releases a Linux desktop supercomputer that's better price/performance wise than anything Wintel is doing and their whole new software stack will only run on WSL2. They aren't porting to Win32. Wow, it may actually be the year of Linux on the Desktop.

I'm a bit surprised by the amount of comments comparing the cost to (often cheap) cloud solutions. Nvidia's value proposition is completely different in my opinion. Say I have a startup in the EU that handles personal data or some company secrets and wants to use an LLM to analyse it (like using RAG). Having that data never leave your basement sure can be worth more than $3000 if performance is not a bottleneck.

There's a market not described here: bioinformatics.

The owner of the market, Illumina, already ships their own bespoke hardware chips in servers called DRAGEN for faster analysis of thousands of genomes. Their main market for this product is in personalised medicine, as genome sequencing in humans is becoming common.

Other companies like Oxford Nanopore use on-board GPUs to call bases (i.e., from raw electric signal coming off the sequencer to A, T, G, C) but it's not working as well as it could due to size and power constraints. I feel like this could be a huge game changer for someone like ONT, especially with cooler stuff like adaptive sequencing.

Other avenues of bioinformatics, such as most day-to-day analysis software, is still very CPU and RAM heavy.

In case you're curious, I googled. It runs this thing called "DGX OS":

"DGX OS 6 Features The following are the key features of DGX OS Release 6:

Based on Ubuntu 22.04 with the latest long-term Linux kernel version 5.15 for the recent hardware and security updates and updates to software packages, such as Python and GCC.

Includes the NVIDIA-optimized Linux kernel, which supports GPU Direct Storage (GDS) without additional patches.

Provides access to all NVIDIA GPU driver branches and CUDA toolkit versions.

Uses the Ubuntu OFED by default with the option to install NVIDIA OFED for additional features.

Supports Secure Boot (requires Ubuntu OFED).

Supports DGX H100/H200."

Nvidia just did what Intel/AMD should have done to threaten CUDA ecosystem - release a "cheap" 128GB local inference appliance/GPU. Well done Nvidia, and it looks bleak for any AI Intel/AMD efforts in the future.

Okay, so this is not a peripheral that you connect to your computer to run specialized tasks, this is a full computer running Linux.

It's a garden hermit. Imagine a future where everyone has one of those(not exactly this version but some future version), it lives with you it learns with you and unlike the cloud based SaaS AI you can teach it things immediately and diverge from the average to your advantage.

>> The IBM Roadrunner was the first supercomputer to reach one petaflop (1 quadrillion floating point operations per second, or FLOPS) on May 25, 2008.

$100M, 2.35MW, 6000 ft^2

>>Designed for AI researchers, data scientists, and students, Project Digits packs Nvidia’s new GB10 Grace Blackwell Superchip, which delivers up to a petaflop of computing performance for prototyping, fine-tuning, and running AI models.

$3000, 1kW, 0.5 ft^2

This looks like a successor to the Nvidia Jetson AGX Orin 64GB Developer Kit:

https://www.okdo.com/wp-content/uploads/2023/03/jetson-agx-o...

I wonder what the specifications are in terms of memory bandwidth and computational capability.

I've followed progress since Moravec's "When will computer hardware match the human brain?" since that came out in 1997. It starts:

>This paper describes how the performance of AI machines tends to improve at the same pace that AI researchers get access to faster hardware. The processing power and memory capacity necessary to match general intellectual performance of the human brain are estimated. Based on extrapolation of past trends and on examination of technologies under development, it is predicted that the required hardware will be available in cheap machines in the 2020s.

and this is about the first personal unit that seems well ahead of his proposed specs. (He estimated 0.1 petaflops. The nvidia thing is "1 petaflop of AI performance at FP4 precision").

(paper https://jetpress.org/volume1/moravec.pdf)

Finally a real ARM workstation from Nvidia! This will be much faster than Apple's offerings for AI work. And at $3000 it is much cheaper than any Mac with 128 GB RAM.

Not much was unveiled but it showed a Blackwell GPU with 1PFLOP of FP4 compute, 128GB unified DDR5X memory, 20 ARM cores, and ConnectX powering two QSFP slots so one can stack multiple of them.

edit: While the title says "personal", Jensen did say this was aimed at startups and similar, so not your living room necessarily.

this image seems to be AI Generated - :-)

https://s3.amazonaws.com/cms.ipressroom.com/219/files/20250/...

Source: https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwe...

$3k for a 128GB standalone is quite favorable pricing considering the next best option at home is going to be a 32GB 5090 at $2k for the card alone, so probably $3k when you’re done building a rig around it.

With more and more personal AI, i think having a truly private device that can run large LLMs (remember: larger is better) is fantastic!

Ideally we can configure things like Apple Intelligence to use this instead of OpenAI and Apple's cloud.

How about “We sell a computer called the tinybox. It comes in two colors + pro.

tinybox red and green are for people looking for a quiet home/office machine. tinybox pro is for people looking for a loud compact rack machine.” [0]

[0] https://tinygrad.org/#tinybox

Is there any effort in local cloud computing? I can't justify $3000 for a fun device. But if all devices (6 phone, 2 iPads, a desktop and 2 laptops) in my home can leverage that for fast LLM, gaming, and photo/video editing, now it makes so much more sense.

Amortized cost is 10 cents per petaflop hour if you run it 5-6 years 24/7. I'm including the cost of electricity.

This is really game changer.

They should make a deal with Valve to turn this into 'superconsole' that can run Half Life 3 (to be announced) :)

I bought a Lamba Labs workstation with a 4090 last year. I guess I’m buying one of these things now because the Lambda workstation just became a relic…

One thing I didn't see mentioned: this would be a good motivation for Nvidia to release "open weights" models.

Just like Mac OS is free when you buy a Mac, having the latest high-quality LLM for free that just happens to run well on this box is a very interesting value-prop. And Nvidia definitely has the compute to make it happen.

What we need is more diversity in the space. Something between the Jetson and this thing, at under $1000 that can run in a LAN to do LLM, STT, TTS, etc. would be an awesome (if niche) device to enable truly local / private AI scenarios for privacy-sensitive folks.

So, a company that doesn't feel like sharing all their secret sauce with Anthropic can run DeepSeek Coder on three of these for $9K, and it should be be more or less the same experience.

Do I understand that right? It seems way to cheap.

If they end up actually shipping this, lots of people will buy these machines to get an AArch64 Linux workstation—even if they are not interested in AI or Nvidia GPUs.

At $3,000, it will be considerably cheaper than alternatives available today (except for SoC boards with extremely poor performance, obviously). I also expect that Nvidia will use its existing distribution channels for this, giving consumers a shot at buying the hardware (without first creating a company and losing consumer protections along the way).

Is this going to make up for the lack of VRAM in the new consumer GPUs?

This is it guys, the most valuable company in the world is selling Linux desktop, after decades of speculating finally the year of Linux desktop is 2025!

Joking aside, personally will buy this workstation in a heartbeat if I have the budget to spare, one in the home and another in the office.

Currently I have an desktop/workstation for AI workloads with similar 128GB RAM that I bought few years back that cost around USD5K without the NVIDIA GPU that I bought earlier for about USD1.5K, with a total of about USD6.5K without a display monitor. This the same price of NeXT workstation (with a monitor) when it's sold back in 1988 without adjusting for inflations (now around USD18K) but it is more than 200 times faster in CPU speed and more than 1000 times RAM capacity than the original 25 MHz CPU and 4 MB RAM, respectively. The later updated version of NeXT has graphic accelerator with 8 MB VRAM, since the workstation has RTX 2080 it is about 1000 times more. I believe the updated NeXT with graphic accelerator is the one that used to develop original Doom software [1].

If NVIDIA can sell the Project Digits Linux desktop at USD3K with similar or more powerful setup configurations, it's going to be a winner and probably can sell by truckloads. It seems to has NeXT workstation vibe to it that used to develop the original WWW and Doom software. Hopefully it will be used to develop many innovative software but now using open source Linux software eco-system not proprietary one.

The latest Linux kernel now has real-time capability for more responsive desktop experience and as saying goes, good things come to those who wait.

[1] NeXT Computer:

https://en.m.wikipedia.org/wiki/NeXT_Computer

I think this is version 1 of what's going to become the new 'PC'.

Future versions will get more capable and smaller, portable.

Can be used to train new types models (not just LLMs).

I assume the GPU can do 3D graphics.

Several of these in a cluster could run multiple powerful models in real time (vision, llm, OCR, 3D navigation, etc).

If successful, millions of such units will be distributed around the world within 1-2 years.

A p2p network of millions of such devices would be a very powerful thing indeed.

Will there be a healthy "personal AI supercomputer" economy to generate demand for this? (NB: spam generators are only a parasite on any digital economy, viable to the extend they don't kill the host).

One can only wish for this, but Nvidia would be going against the decades-long trend to emaciate local computing in favor of concentrating all compute on somebody else's linux (aka: cloud).

Can one game on it?

If one can skip buying gaming rig with a 5090 with its likely absurd price then this 3k becomes a lot easier for dual use hobbyists to swallow

Edit 5090 is 2k

Smart move for Nvidia to subsidise their ARM CPU and platform business by selling a big GPU packet with a CPU that most users don't really care about. Even if the margin is less than selling the raw GPU power would be (which I doubt), it'll look good on the shareholders conference if other business segments go up steep, too.

Little by little, we're getting an answer to the question: "What kind of investment does an outrageous influx of capitalization spur?" One might think it would be an AI silicon-moat, and it might yet be some of that.

But it's clear that everyone's favorite goal is keretsuification. If you're looking for abnormal profits, you can't do better than to add a letter to FAANG. Nvidia already got into the cloud business, and now it's making workstations.

The era of specialists doing specialist things is not really behind us. They're just not making automatic money, nor most of it. Nvidia excelled in that pool, but it too can't wait to leave it. It knows it can always fail as a specialist, but not as a kereitsu.

This seems surprisingly cheap for what you get! Excited to see what people cook with this

It's based on Mediatek CPU cores, so I am really pessimistic about their open source support...

I'm bracing for a whole new era of unsufferable binary blobs for Linux users, and my condolences if you have a non-ultramainstream distro.

I'm buying one. It's cheaper then my 4090RTX+192GB of ram for more performance and model traning headroom. It's also probably a beast for data science workloads.

I am pretty sure memory bandwidth will be low it doesn't eat up their enterprise lineup. If we are luck we might get 512GB/S this is still half of 4090

With a bit of luck it'll mean some of the Jetson series will get cheaper.

While I'm quite the "AI" sceptic I think it might be interesting to have a node in my home network capable of a bit of this and that in this area, some text-to-speech, speech-to-text, object identification, which to be decent needs a bit more than the usual IoT- and ESP-chips can manage.

The monitor is AI generated in the product photo.... Nvidia please

https://s3.amazonaws.com/cms.ipressroom.com/219/files/20250/...

Running large language models locally could be the next major shift in personal computing, similar to how GPUs transformed gaming. The privacy implications of not sending data to cloud services are significant.

I highly doubt it's half or ever quarter of GB200, unless they have hidden water cooling or something outside. GB200 is 1200 Watts. Digits doesn't look like it would be above 200W, and cooling 200W would be impressive.

> It’s a cloud computing platform that sits on your desk …

This goes against every definition of cloud that I know off. Again proving that 'cloud' means whatever you want it to mean.

What are the CPU specs? Idk about the GPU but a really fast ARM cpu and a ton of ram and it already runs Linux?!! If it’s competitive with the M chips from Apple this might be my next box.

Finally!

First product that directly competes on price with Macs for local inferencing of large LLMs (higher RAM). And likely outperforms them substantially.

Definitely will upgrade my home LLM server if specs bear out.

Does this also answer the question "What am I supposed to do with my old 4090 and my old 3090 once I get a 5090?" ie can we attach them as PCIe hardware to Digits?

Nvidia could potentially bring us all the year of Linux Desktop.

$3000? The GB10 inside it seems to be a half of GB200 which is like $60K. One can wonder about availability at those $3K.

Would love to see something like this with an ATX form factor, socketed GPU, socketed GPU and non-soldered RAM.

$3000 seems incredibly good value

Newbie question: can you connect a RTX 5090 to the GB10 Digits computer?

Of course no chance of this with x86 because of market segmentation.

Feels like data center AI HW demand peak is over now that these things are trickling down to consumers and they are diversifying customers. Also going lower than expected on gaming HW, seems like they have enough fab capacity.

Do I need a "personal AI supercomputer"?

Anyone know if it can run training/fine-tuning and not just 4-bit inference? Does it support mixed precision training with either BF16 or FP16?

Honestly surprised at how affordable this is, I was expecting $5-6k as I scanned the opening paragraphs.

Pricing seems off!

Welcome to tomorrow's "personal" computer, a single unmodifiable SoC with closed source software stack.

...but will it run DOOM?

It costs the equivalent of 2 years of cloud GPU H100s at current prices.

Edit: Sorry fucked up my math. I wanted to do 40x52x4, $4/hr being the cloud compute price but that us actually $8300, so it is actually equivalent to about 4.5 months of cloud compute. 40 hours because I presume that this will only be used for prototyping and debugging, i.e during office hours.

Would consider at a lower price of $500 USD, way too expensive for what it brings.