At NVIDIA GTC 2025, NVIDIA launched what we expect to be the ultra-popular platform in the GB10 with small options like the NVIDIA DGX Spark and ASUS Ascent GX10. There was another, much larger platform released, the new NVIDIA DGX Station GB300 edition. Despite sporting a NVIDIA Grace Arm-based CPU and a NVIDIA Blackwell GPU, the platform requires installing another GPU to get video output.
NVIDIA DGX Station GB300 Edition Launched Without a GPU
Instead of selling the DGX Station as a NVIDIA-only product, the company has made a standard(ish) motherboard and plans to sell it through OEMs like Dell and HP.
At the heart of the system is the NVIDIA GB300.
The NVIDIA Grace is a 72-core Arm CPU with PCIe Gen5. It is also adopting Micron SOCAMM Memory in all of the systems we have seen. NVIDIA says this is good for up to 496GB of LPDDR5X with 396GB/s of memory bandwidth. That is OK, but x86 CPUs already have higher memory bandwidth, especially with 12-channel memory designs.
There is a M.2 slot nexto the Grace CPU and the Blackwell B300 GPU. The B300 is listed as the 288GB HBM3e at 8TB/s.
On the front edge of the motherboard, there are two more M.2 slots, the BMC, and a black chip. The chip being covered here is an FPGA for the platform.
In the PCIe slots of both the Dell and HP systems we saw, there were NVIDIA RTX 4000 Ada GPUs. This is for a strange reason. Despite the B300, there are no onboard video outputs, so the platform needs a GPU to be a workstation.
The two demo systems we saw inside of had ASPEED AST2600 BMC modules installed.
Perhaps the most interesting was the top portion. There are two Realtek RTL8111 NICs onboard, but then the big one, a NVIDIA ConnectX-8. NVIDIA says this is good for up to 800Gbps of networking.
Along the top side of the motherboard we also get standard power connectors, making this seemingly easy to integrated for NVIDIA’s partners.
Final Words
The title of this is there to help folks think about calling the B300/ GB300 a “GPU” since there is an additional card required for graphics output. Aside from that, this looks like an awesome Arm-based workstation for those doing AI work. To be clear, the NVIDIA Grace CPU at 72 cores does not have the performance of modern Intel or AMD server class (or Threadripper workstation class) processors. The point of Grace is really to move data to support the GPUs and provide connectivity. In 2022, we called an Arm system with NVIDIA A100’s The Most Important Server of 2022. Now, we are seeing NVIDIA push into building big systems with its GB300 architecture, and we finally have a development platform. If you are an AI researcher this is going to be a top option over simply adding PCIe GPUs to another platform. Hopefully we get to see these in the market soon.
Maybe the big question will end up being whether these are used as workstations or servers given their specs.
Spell check
An overview of the NVIDIA Spark GB10 128GB and Workstation GB300 compared to the other NVIDIA GPU’s was not given yet by NVIDIA. Typical marketing intransparencies. No real product comparison possible at the moment..
How many tensor cores, cuda cores on the NVIDIA Spark GB10 128GB and Workstation GB300?
What is the memory bandwidth in practice? How many tokens per sec for deepseekR1:70b at Q4?
everybody gansta till i install an AMD gpu to that…
Interesting, they are using 3 12VHPWR connectors for supply on the top of the board.
Can’t wait for some in-depth analysis of the board. I want to see if it suffers from the same design issues as the gaming GPUs.
“a NVIDIA” lol.
“spell check March 28, 2025 At 12:22 pm
“a NVIDIA” lol.”
Imagine being that troll that doesn’t realize that NVIDIA is not an acronym. Even then it’s not even proper to do it. Yeppers 1000IQ move.
Shame. I really wanted to see Crysis run on these GPUs. Not that anyone would buy them for that purpose, but it would have been really cool.
Xander, it has nothing to do with whether or not NVIDIA is an acronym. From the Purdue University Online Writing Lab: “if the first letter makes a vowel-type sound, you use ‘an’.” Because “Nvidia Corporation” is pronounced “(/ɛnˈvɪdiə/ en-VID-ee-ə),” that is to say, starting with the same vowel sound as “envy” or “environment,” the proper article is “an.” That this information is no longer common knowledge speaks volumes about the declining quality of education.