News Nvidia's Project Digits desktop AI supercomputer fits in the palm of your hand — $3,000 to bring 1 PFLOPS of performance home

[Admin]

Nvidia announced the "world's smallest AI supercomputer" at CES with Project Digits, a 1 PFLOPS machine to handle the entire Nvidia enterprise software stack.

Nvidia's Project Digits desktop AI supercomputer fits in the palm of your hand — $3,000 to bring 1 PFLOPS of performance home : Read more

 

[bit_user]

Dear Tom's authors & editors,

Please start putting quotes around Nvidia's use of the term "supercomputer", whenever they're is using it to describe something like an embedded computer or desktop mini. The quoted performance is only about as fast as the newly-announced RTX 5070, which is slated to cost just $549. So, it basically sounds like an overpriced ARM-based mid-range gaming PC. Not even a super gaming PC.

A real supercomputer consists of multiple racks and is comprised of thousands of GPUs and at least hundreds of CPUs. If this is a supercomputer, then so is just about every gaming PC that has a RTX 5070 or better. Their abuse of the term is rendering it meaningless.


P.S. the main thing that makes this special is the memory capacity, not the PFLOPS. If they'd sell dGPUs with that much memory, they'd be a much better option for inferencing such models.

 

[Loadedaxe]

wow, the New maxed out Mac Mini is useless now.

Its sarcasm.....put down the pitchforks! 😉

 

[bit_user]

wow, the New maxed out Mac Mini is useless now.

Its sarcasm.....put down the pitchforks! 😉

I've got to say the core configuration is pretty decent, though. 10x of ARM latest & greatest P-cores + 10x of their latest mid cores.

https://www.anandtech.com/show/2139...5-a725-and-a520-arm-v9-2-redefined-for-3nm-/2
​

Is it going to beat 8x Lion Cove + 16x Skymont? Hard to say, but I think it should certainly be in the same league.

 

[jcihanj]

This Project DIGITS GB10 might be on the wish list of all the AI cloud providers: Google (ASTRA Augmented Reality), Meta (Meta Ray Ban Augmented Reality) and Apple (Apple Visioin Pro ), and MSFT's Hololens. The reason is GB10 is powerful enough to embody the front end muti-modal AI interface that handles audio and visual queries from end users. The 1 PFLOPs (1000 TOPS) of processing can simultaneously support multiple high performance FP8 versions of audio, visual LLM models, which then enables highly responsive AI-driven communication interface required to realize AR applications. These vision-centric multimodal queries are first processed by the GB10 at the edge, where the users reside; and if a solution is not available locally, a query is then forwarded to on-prem enterprise AI servers, or to the cloud-based AI servers, for the ultimate solution search. Thus this GB10 could enable AR to become one of the killer apps of 2025.