The NVIDIA DGX Spark (Spark Box) is a compact desktop‑form‑factor personal AI supercomputer built on the Grace Blackwell GB10 Superchip, integrating a 20‑core ARM Grace CPU (10×Cortex‑X925 +10×Cortex‑A725) and Blackwell‑generation GPU on one packageNVIDIA. It features 5th‑Gen Tensor Cores and 4th‑Gen Ray Tracing Cores, delivering up to 1 PFLOPS FP4 AI throughput for agentic AI workloads, large‑model inference and small‑scale model fine‑tuning.
Equipped with 128 GB coherent unified LPDDR5X system memory (273 GB/s bandwidth), memory is shared between CPU and GPU, eliminating traditional CPU‑GPU data transfer bottlenecks. It ships with a 4 TB self‑encrypted NVMe M.2 SSD for storing model weights and local datasetsNVIDIA.
Advanced high‑speed networking includes ConnectX‑7 200 Gbps smart NIC plus 10 GbE RJ45 port, Wi‑Fi 7 and Bluetooth 5.4. Multiple DGX Spark units can be clustered together to expand total memory and compute capacity for models up to 405 B parameters in dual‑node configurationNVIDIA. The unit has four USB‑C ports and HDMI 2.1a video output. Pre‑loaded with DGX‑OS, full NVIDIA CUDA software stack, NemoClaw agent framework out‑of‑the‑box. Total system power is 240 W with GB10 superchip TDP at 140 W, delivering enterprise‑grade AI capability within a small desktop footprint.
It is ideal for AI research teams, agent developers and data scientists running fully‑private local AI workloads without public‑cloud tokens and latency.
Official Model: NVIDIA DGX Spark (Spark Box)
Superchip: NVIDIA GB10 Grace Blackwell
CPU: 20‑core ARM (10 × Cortex‑X925 + 10 × Cortex‑A725)
GPU: Blackwell GPU, 5th‑Gen Tensor Cores, 4th‑Gen RT Cores
AI Compute: Up to 1 PFLOPS FP4
System Memory: 128 GB Coherent Unified LPDDR5X‑8000; Bandwidth:273 GB/s
Internal Storage: 4 TB Self‑Encrypted NVMe M.2 SSD
Networking: ConnectX‑7 200 Gbps NIC, 10 GbE RJ45, Wi‑Fi 7, Bluetooth 5.4
I/O: 4 × USB‑C, HDMI 2.1a video output
OS: Pre‑installed DGX‑OS, full NVIDIA CUDA software stack
TDP: GB10 Chip 140W; Total System Power:240W
Supported Workloads: Local AI‑Agent development, offline LLM inference, lightweight fine‑tuning, multi‑node cluster capable
Max supported single‑node model size: Up to 200 B‑parameter LLMs
Note: 64GB memory variant is also available as an alternative SKU.