Thanks. This has to be one of the subtlest marketing/nomenclature distinctions I've ever seen.
The one-word difference between [1] the "GB300 Grace Blackwell Ultra Superchip" used for a building block for
GB300 NVL72 rack-scale servers and [2] the "GB300 Grace Blackwell Ultra
Desktop Superchip" used for the
DGX Station is quite large: the server has quadruple the GPU and double the CPU of the workstation.
First, "SuperChip" shouldn't be though of as a 'chip' or a 'chip package'. Pragmatically it is really a logicboard ( printed circuit board; PCB). One thing that Nvidia is trying to do is claim they are not in the systems maker busines. That they are still just a major component vendor, that isn't trying to compete with the major system sellers (HPE, Dell, Supermicro , etc. )
The Nvidia C2C-Link is essentially a link via a PCB to to 'Chip' packages. (C2C .... chip to chip ).
Chip Interconnect Technology
www.nvidia.com
Some terminology would tag it as "multiple chip module" , but "SuperChip" sounds snazzier.
This NVL72 "Superchip" is more akin to the size of a main logic board.
The assertion that is some kind of "chip" where there are screws holding down subcomponents , networking sockets, cylindrical capacitors , etc. it is rally a 'board" or at best "module". That is huge. ( and also power consuming past normal wall socket circuit capacity; at least USA codes. )
The "GB 300" prefix is really about there just being a "Grace" package in the mix along with some "Blackwell" silicon. A guide from back on GB200 era (like that was long ago.
).
A few days ago, xAI announced Grok 3 and it’s easily the best LLM I’ve used. Good vibes, great quality, fast token generation, and happened to be trained on the most compute for any model (200,000 H100s). During the stream, Elon Musk said xAI is already working on Grok 4’s data center, which will be
creativestrategies.com
The problem is that logic board is really too big for a personal workstation, if going to normal stuff that goes on a workstation (e.g., local storage, a couple of general PCI-e slots. , regular workstation I/O sockets). The above is more so designed to put two 'nodes' on teh same board. (where a node is Grace+ 2Blackwells. ). It is a building block aimed at a different scale. The "GB 300" is really only about the basic node structure there, no the whole board.
The workstation doesn't have two nodes. Also trims back the Blackwell parts also as both cost and space saving.
However, has enough in common with the larger system so local development work should scale straightforwardly to a cluster. These 'dev' boxes don't have to sit in a datacenter with high end HVAC.
Nvidia refers to the GB300 memory as (up to) "784GB of large coherent memory," (288GB HBM3e + 496GB LPDDR5X) so I assume the purpose-built DGX OS must see it as something like Apple's unified memory architecture. Is that a safe assumption?
Coherent is more suggestive that have a 'flat address' space. ( technically coherent means the changes made to memory are seen by everyone who might have a copy. But for everyone to know is shared they all need a common label; hence common addresses. ) That doesn't necessarily mean "uniform". Apple's 'Unified' has a substantive 'uniform' assumption built into it.
With regard to the Mac Pro, I think I'd expect Apple to blend the two. That's beauty of Ultra = 2x Max. Package an Ultra with a building block created for the AI servers and voila, best of both worlds!
Not much "chiplet" beauty there. Once you have coupled to 'Max' dies to each other , you have basically used up all of the highspeed , low latency connections ( at least if competing with C2C-link type of performance) . You might be able to take one Max and couple it to something else that wasn't laptop optimized , but that is it.
Nvidia's data center GPUs have both C2C-Link and NVLink connectors. So they are purpose built to "scale up" on the same logic board. (e.g., the two nodes on the NVL72 board can both talk to each other in additional to other GPUs on other boards in the same cabinet.
What are the odds? I don't know, but not super high. Still, if you watch that Nvidia keynote, where he introduces Vera Rubin and its second generation, Rubin Ultra, it feels like Apple needs to make some kind of statement, hardware-wise.
Apple's AI solution doesn't have to outperform Nvidia's in raw compute. It just has to be cheaper ( more affordable within the power consumption parameters Apple wants to constraint it too.) Apple isn't going to sell the hardware to anyone else so what meaningful 'statement' could they possibly make? The point is to not write billion dollar checks to Google** or Nvidia ( or OpenAI or Microsoft or etc. ). Apple only needs a 'statement' if trying to make other folks change their Nvidia data center buying habits. There is little indication Apple is chasing that at all.
Extremely likely Apple is going to be looking a solution that gets better Perf/Watt than Nvidia's does. (e.g., C2C-link costs more in power than UltraFusion does. )
If Apple 'sells' anything it might be for the AI cloud service , but isn't a 'hardware' sale. So far Apple saying it is all free... so not much 'statement' making there either. ( and if 'free' then all the more likely that the power consumption bill will matter at least as much as benchmark bragging rights. )
** Some reports are that Apple is using Google Cloud services and Tensors to do substantive parts of the training. Apple's AI server chipls likely will focus on 'inference' for Apple's service cloud , if they are reasonably decent at training also that would likely be a candidate to offload from outsourcing also (if cheaper).
