NVIDIA GB10 Grace Blackwell Superchip

The desktop AI supercomputer.

DGX Spark 128GB puts a petaFLOP of AI performance and 128 GB of unified memory on your desk, enough to run models up to 200 billion parameters locally and privately.

NVIDIA DGX Spark in its gold chassis in front of AI workload tiles
1 PFLOPFP4 AI performance
128 GBUnified LPDDR5x memory
20-coreArm CPU
200BParameter models, locally

Overview

A complete platform for local AI.

Powered by the NVIDIA GB10 Grace Blackwell Superchip, DGX Spark combines a 20-core Arm CPU and a Blackwell GPU that share one coherent pool of memory. It ships with the NVIDIA AI software stack, so agents and large models run on your own hardware instead of rented cloud tokens.

  • Grace Blackwell architecture with CPU and GPU sharing 128 GB of unified memory.
  • Palm-sized 150 mm chassis that weighs just 1.2 kg and runs from a 240 W adapter.
  • ConnectX-7 networking links up to four units into a single desktop cluster.
  • Preinstalled DGX OS and NVIDIA AI stack, plus NVIDIA Sync for remote access.
DGX Spark on a desk next to a laptop running an AI agent

Workloads

Accelerate every AI workload.

Built for AI developers, researchers and data scientists who need data-center architecture at the desk.

Agentic AI

Build, run and manage always-on AI agents privately from your desk, from single-agent prototypes to multi-agent workflows.

Prototyping

Develop, test and validate models and AI-augmented applications before migrating to DGX Cloud or the data center.

Fine-tuning

Customize pre-trained models of up to 70 billion parameters using the full 128 GB of unified memory.

Inference

Run and validate models up to 200 billion parameters with FP4 support on fifth-generation Tensor Cores.

Data science

Accelerate large analytics and machine-learning pipelines with 1 PFLOP of parallel throughput.

Edge applications

Develop robotics, smart-city and vision solutions with NVIDIA Isaac, Metropolis and Holoscan.

Specifications

DGX Spark 128GB tech specs.

Compute

Superchip
NVIDIA GB10 Grace Blackwell
AI performance
Up to 1 PFLOP (1,000 TOPS) at FP4 with sparsity
CPU
20-core Arm (10× Cortex-X925 + 10× Cortex-A725)
GPU
Blackwell architecture, 5th-gen Tensor Cores, 4th-gen RT Cores
CUDA cores
6,144
Video
1× NVENC, 1× NVDEC

Memory & Storage

System memory
128 GB LPDDR5x, coherent unified memory
Memory interface
256-bit, 273 GB/s bandwidth
Storage
1 TB or 4 TB NVMe M.2, self-encrypting

Connectivity

Ethernet
1× RJ-45 10 GbE
High-speed networking
NVIDIA ConnectX-7 SmartNIC, 2× QSFP
Wireless
Wi-Fi 7, Bluetooth 5.4
USB
4× USB Type-C (one for power delivery)
Display
1× HDMI 2.1a with multichannel audio

Physical & Power

Dimensions
150 × 150 × 50.5 mm
Weight
1.2 kg (2.6 lbs)
Power supply
240 W external adapter (included)
SoC TDP
140 W
Operating temperature
5 °C – 30 °C (41 °F – 86 °F)
Software
NVIDIA DGX OS with the NVIDIA AI software stack
Rear panel of DGX Spark: power button, power USB-C, three USB-C, HDMI, Ethernet and two QSFP ports
Rear panel: on/off, power, USB-C, HDMI, 10 GbE Ethernet and dual QSFP (ConnectX-7).

Scale

Link systems for frontier-class models.

Connect up to four DGX Spark units with ConnectX-7 to pool memory for larger models and faster inference.

1× DGX Spark

128 GB Up to 200B parameters

2× DGX Spark

256 GB Up to 400B parameters

4× DGX Spark

512 GB Up to 700B parameters

Get started

Your own AI supercomputer.

Explore step-by-step playbooks and find where to buy on NVIDIA's official pages.