During the NVIDIA GTC conference in San Jose, CA, the GPU giant announced two small supercomputers: the DGX Spark and DGX Station. Both supercomputers use the NVIDIA Blackwell Ultra platform and are targeted to developers, researchers, data scientists, and students training, running inference, and deploying large language models.
“AI has transformed every layer of the computing stack. It stands to reason a new class of computers would emerge — designed for AI-native developers and to run AI-native applications,” said NVIDIA CEO and cofounder Jensen Huang in a press release. “With these new DGX personal AI computers, AI can span from cloud services to desktop and edge applications.”
DGX Spark has 784 GB of memory in a small package
NVIDIA claimed the DGX Spark, previously known as Project Digits, is the world’s smallest supercomputer. The GB10 Grace Blackwell Superchip inside includes a Blackwell GPU that can perform 1,000 trillion operations per second of AI computing. NVIDIA NVLink-C2C interconnect technology hooks a CPU and the GPU together at five times the bandwidth of fifth-generation PCIe, NVIDIA said.
From DGX Spark, users can export AI models to DGX Cloud or any accelerated cloud or data center infrastructure.
SEE: Microsoft accidentally removed its Copilot AI assistant from some devices – and some users are glad to hear it.
DGX Spark can run NVIDIA’s recently-announced GR00T N1 robotics foundation model, and the Cosmos world generation system that teaches AI-powered robots about the physical world.
A waitlist is open now for the DGX Spark, which will cost $3,000. The DGX Spark will be produced by manufacturers including ASUS, Dell, and HP.
DGX Station is built for desktop AI development
The updated edition of the DGX Station will be available later this year from NVIDIA’s manufacturing partners, including ASUS, BOXX, Dell, HP, Lambda, and Supermicro. The price has not been disclosed.
With the current-gen DGX Station, users receive 784 GB of coherent memory space, suitable for training and inferencing for large AI workloads. Inside is the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, the Blackwell Ultra GPU, and the NVIDIA Grace CPU, all connected by the NVLink-C2C.
Inside the supercomputer also sits an NVIDIA ConnectX-8 SuperNIC, which supports networking at up to 800Gb/s. It can link several DGX stations together for massive workloads or for making network-accelerated data transfers.
Plus, the DGX Station interoperates with the NVIDIA CUDA-X AI platform for desktop AI development, the NVIDIA NIM microservices, and the NVIDIA AI Enterprise software platform.