In the course of the NVIDIA GTC convention in San Jose, CA, the GPU large introduced two small supercomputers: the DGX Spark and DGX Station. Each supercomputers use the NVIDIA Blackwell Extremely platform and are focused to builders, researchers, knowledge scientists, and college students coaching, working inference, and deploying massive language fashions.
“AI has reworked each layer of the computing stack. It stands to motive a brand new class of computer systems would emerge — designed for AI-native builders and to run AI-native functions,” stated NVIDIA CEO and cofounder Jensen Huang in a press launch. “With these new DGX private AI computer systems, AI can span from cloud providers to desktop and edge functions.”
DGX Spark has 784 GB of reminiscence in a small bundle
NVIDIA claimed the DGX Spark, beforehand referred to as Undertaking Digits, is the world’s smallest supercomputer. The GB10 Grace Blackwell Superchip inside features a Blackwell GPU that may carry out 1,000 trillion operations per second of AI computing. NVIDIA NVLink-C2C interconnect expertise hooks a CPU and the GPU collectively at 5 occasions the bandwidth of fifth-generation PCIe, NVIDIA stated.
From DGX Spark, customers can export AI fashions to DGX Cloud or any accelerated cloud or knowledge middle infrastructure.
SEE: Microsoft unintentionally eliminated its Copilot AI assistant from some units – and a few customers are glad to listen to it.
DGX Spark can run NVIDIA’s recently-announced GR00T N1 robotics basis mannequin, and the Cosmos world technology system that teaches AI-powered robots concerning the bodily world.
A waitlist is open now for the DGX Spark, which is able to price $3,000. The DGX Spark will probably be produced by producers together with ASUS, Dell, and HP.
DGX Station is constructed for desktop AI improvement
The up to date version of the DGX Station will probably be accessible later this yr from NVIDIA’s manufacturing companions, together with ASUS, BOXX, Dell, HP, Lambda, and Supermicro. The value has not been disclosed.
With the current-gen DGX Station, customers obtain 784 GB of coherent reminiscence house, appropriate for coaching and inferencing for giant AI workloads. Inside is the NVIDIA GB300 Grace Blackwell Extremely Desktop Superchip, the Blackwell Extremely GPU, and the NVIDIA Grace CPU, all related by the NVLink-C2C.
Contained in the supercomputer additionally sits an NVIDIA ConnectX-8 SuperNIC, which helps networking at as much as 800Gb/s. It could possibly hyperlink a number of DGX stations collectively for large workloads or for making network-accelerated knowledge transfers.
Plus, the DGX Station interoperates with the NVIDIA CUDA-X AI platform for desktop AI improvement, the NVIDIA NIM microservices, and the NVIDIA AI Enterprise software program platform.
