Nvidia has announced the launch of what calls a new line of «Ia personal supercomputers» promoted by the Grace Blackwell chips platform, an architecture that already uses on its latest generation dedicated graphics cards.
The personal supercomputers of AI have been one of the novelties of the Keynote that the CEO of Nvidia, Jensen Huang, offered last night at his GTC 2025 conference. Taking into account the extraordinary potential that Nvidia has reached in hardware for artificial intelligence, it was expected that AI and deep learning were a central point of the conference.
“This is the AI era computer”Huang assured during the presentation of the new machines that will allow users to protect, improve and execute models of various sizes on the edge. «Thus computers should be seen and thus they will work in the future. And now we have a complete range for companies, from the smallest to the workstations ».
“AI agents will be everywhere,” Huang continued. «Its operation, that of companies and ours will be fundamentally different. That is why we need a new computers line. And this is the solution, ”said the executive when presenting his new personal supercomputers of AI.
DGX Spark
Known in its development as ‘Project Digits’, it is a machine that bets on a surprisingly compact size of 150 X 150 x 50.5mm, even smaller than a typical mini-PC. Includes the new chip Grace GB10 Created in collaboration with Mediatek and includes a Blackwell GPU connected to a 20 -core CPU Grace.
Within the housing, the chips are connected to a 128 GBytes memory set and has capacity for up to 4 tb of storage with solid state units. It has support for Wi-Fi 7 and Bluetooth 5.3 and a set of ports that includes four USB4 (40 Gbps) ports
1 LAN of 10 GBE (Intelligent Nic Connectx-7) and an HDMI 2.1 for external screens connection. The total energy consumption of the system is 170 watts.
Spark focuses on IA researchers, data scientists, developers and students, this solution to a petaflop (up to 1000 billion operations per second computing AI) to create prototypes, adjust and execute AI models.
Nvidia ensures that a single unit can execute models with a size of up to 200 billion parameters. The parameters correspond approximately to the problem solving skills of a model, and the most parameters models usually work better than those with less parameters. You can even connect two machines to each other to execute models of up to 405 billion parameters, if a job requires it.
The DGX SPARK works with the DGX operating system based on NVIDIA Linux and can function independently or connected to a main PC that runs Windows or MacOS. It is now available to book and with a Price in the $ 3,000 environments.
DGX Station
The second of the personal supercomputers of AI is destined for researchers and data scientists who need even more the processing power than the one offered by the DGX SPARK. Similar in size to a desktop PC in tower format, use a Grace Blackwell Ultra GB300 chip as the main engine.
The chip includes a 72 -centers Neo Tape V2 and can climb up to 784 GB of unified system memory (288 GB of GPU + 496 GB of LPDDR5X for the CPU). It has a wide connectivity, highlighting the Network with a LEPTX-8 Supernic with speeds of up to 800 GB/s.
Nvidia has not yet announced the price of the DGX Station, although the company states that the computer will be available at the end of this year. It will work with the NVIDIA DGX OS operating system and it is expected that hardware partners such as ASUS, Boxx, Dell, HP, Lambda or Supermicro develop their own versions of the system.
