EN
NVIDIA DGX Spark AI Supercomputer — GB10 Grace Blackwell, 128GB Unified Memory, 4TB NVMe, 1 PFLOP FP4
Superchip:NVIDIA GB10 Grace Blackwell Superchip
CPU: 20-core Arm CPU (10× Cortex-X925 + 10× Cortex-A725)
GPU:Blackwell GPU with 5th-generation Tensor Cores, up to 1 PFLOP at FP4 precision
Memory:128GB coherent unified memory, 256-bit LPDDR5x at 273 GB/s
- Overview
- Recommended Products
Key Features
NVIDIA GB10 Grace Blackwell Superchip — 20-core Arm CPU (10× Cortex-X925 + 10× Cortex-A725) paired with a Blackwell GPU delivering up to 1 PFLOP at FP4 precision
128GB Coherent Unified Memory — 256-bit LPDDR5x at 273 GB/s; run, fine-tune and serve models up to 200 billion parameters locally
4TB NVMe M.2 Storage with self-encryption, preinstalled with NVIDIA DGX OS and the full NVIDIA AI software stack
Scale Out When You Need More — NVIDIA ConnectX-7 networking (200 Gb/s, 2× QSFP) plus 10 GbE and Wi-Fi 7; cluster up to 4 units to work with models up to 700 billion parameters
True Desktop Footprint — 150 × 150 × 50.5 mm, 1.2 kg, powered by a single 240W external supply
Description
NVIDIA DGX Spark brings the Grace Blackwell architecture to the desktop. Built for AI developers, researchers and data scientists, it lets you prototype, fine-tune and run inference on large language models without sending data to the cloud.
The GB10 Superchip combines a 20-core Arm CPU with a Blackwell GPU with 5th-generation Tensor Cores, connected through NVLink-C2C for CPU+GPU coherent memory. With 128GB of unified memory you can work with models up to 200 billion parameters on a single unit, and scale to 405B across two units — or up to 700B across four — using built-in ConnectX-7 networking.
DGX Spark ships with NVIDIA DGX OS and a preinstalled AI software stack, with access to PyTorch, TensorRT-LLM, Jupyter, Ollama, NVIDIA NIM and NVIDIA Blueprints. Work prototyped locally moves directly to DGX Cloud or any NVIDIA-accelerated data center.
Target applications: local LLM inference · model fine-tuning · AI agent development · data science · edge AI, robotics and computer vision prototyping.
Technical Specifications
Product Type |
NVIDIA DGX Spark AI supercomputer (desktop Grace Blackwell system) |
Superchip |
NVIDIA GB10 Grace Blackwell Superchip |
CPU |
20-core Arm CPU (10× Cortex-X925 + 10× Cortex-A725) |
GPU |
Blackwell GPU with 5th-generation Tensor Cores, up to 1 PFLOP at FP4 precision |
CPU-GPU Interconnect |
NVLink-C2C coherent memory link |
Memory |
128GB coherent unified memory, 256-bit LPDDR5x at 273 GB/s |
Model Size |
Up to 200 billion parameters on one unit; 405 billion across two units; up to 700 billion across four units |
Storage |
4TB NVMe M.2 with self-encryption |
Networking |
NVIDIA ConnectX-7 (200 Gb/s, 2× QSFP), 10 GbE and Wi-Fi 7 |
Operating System |
NVIDIA DGX OS |
AI Software Stack |
PyTorch, TensorRT-LLM, Jupyter, Ollama, NVIDIA NIM and NVIDIA Blueprints, preinstalled |
Deployment Path |
Move locally prototyped work to NVIDIA DGX Cloud or any NVIDIA-accelerated data center |
Power Supply |
Single 240W external supply |
Dimensions |
150 × 150 × 50.5 mm |
Weight |
1.2 kg |
Target Applications |
Local LLM inference, model fine-tuning, AI agent development, data science, edge AI, robotics and computer vision prototyping |
