GB10 Grace Blackwell Superchip · 20-Core ARM · ConnectX-7 · Wi-Fi 7 · Desktop
NVIDIA GB10 Grace Blackwell
1 PFLOPS (FP4)
Blackwell, 5th-Gen Tensor Cores
20-Core ARM (10x X925 + 10x A725)
128 GB LPDDR5x Unified
Up to 273 GB/s
4 TB NVMe M.2 (Self-Encrypting)
ConnectX-7 200Gb/s + 10GbE RJ-45
Wi-Fi 7 + Bluetooth 5.4
1x HDMI 2.1a
4x USB Type-C
240W
150 x 150 x 50.5 mm
1.2 kg
NVIDIA DGX OS (Ubuntu Linux)
200B params (405B dual-Spark)
Available
Quote
The NVIDIA DGX Spark is the world's smallest AI supercomputer, bringing petaflop-class AI compute directly to the developer desktop. Powered by the NVIDIA GB10 Grace Blackwell Superchip, it delivers up to 1 PFLOPS of FP4 AI performance with 128GB of coherent unified system memory. This architecture enables running inference on models up to 200 billion parameters and fine-tuning models up to 70 billion parameters — all locally, without cloud dependencies.
At its core, DGX Spark combines a 20-core Arm CPU (10 Cortex-X925 + 10 Cortex-A725) with a Blackwell GPU featuring 5th-generation Tensor Cores and 4th-generation RT Cores. NVLink-C2C interconnect provides 5x the bandwidth of PCIe Gen 5 between CPU and GPU. The system includes a ConnectX-7 Smart NIC delivering 200 Gb/s networking for high-speed data access and clustering. Dual-Spark configurations support models up to 405 billion parameters.
DGX Spark comes preinstalled with NVIDIA DGX OS (Ubuntu Linux) and the full NVIDIA AI software stack — CUDA libraries, NVIDIA NIM microservices, and optimized frameworks. Developers can immediately begin prototyping with PyTorch, Jupyter, Ollama, and other popular tools. Available through NVIDIA and global partners including ASUS, Dell, HP, Lenovo, and GIGABYTE.