NVIDIA DGX Spark

DGX personal AI computer, designed to build
and run AI.
Desktop AI Compute Demands
The increasing size and complexity of generative AI models is making development
efforts on local systems challenging. Prototyping, tuning, and inferencing large
models locally requires large amounts of memory and significant compute performance.
As enterprises, software providers, government agencies, startups, and researchers
staff up AI efforts, the need for AI compute resources continues to grow.
200B Parameter Models on Your Desk
NVIDIA DGX™ Spark is part of a new class of computers designed from the ground
up to build and run AI. Powered by the NVIDIA GB10 Grace Blackwell Superchip and
based on the NVIDIA Grace Blackwell architecture, NVIDIA DGX Spark delivers up to
1000 TOPS of AI performance to power large AI workloads. With 128 GB of unified
system memory, developers can experiment, fine-tune, or inference models of up to
200B parameters. Plus, NVIDIA ConnectX™ networking can connect two NVIDIA DGX
Spark supercomputers to enable inference on models up to 405B parameters.
To give developers a familiar experience, NVIDIA DGX Spark mirrors the same
software architecture that powers industrial-strength AI factories. Using the NVIDIA
DGX OS with Ubuntu Linux and preconfigured with the latest NVIDIA AI software
stack, along with developer program access to NVIDIA NIM™ and NVIDIA Blueprints,
developers can hit the ground running using common tools such as Pytorch,
Jupyter, and Ollama to prototype, fine-tune, and inference on NVIDIA DGX Spark
and seamlessly deploy in the data center or cloud.
By delivering massive performance and capabilities in a compact package, NVIDIA
DGX Spark lets developers, researchers, data scientists, and students continue to
push the boundaries of generative AI.
Built on NVIDIA Grace Blackwell
At the heart of NVIDIA DGX Spark is the new NVIDIA GB10 Grace Blackwell Superchip
based on the NVIDIA Grace Blackwell architecture optimized for a desktop form
factor. GB10 features a powerful NVIDIA Blackwell GPU with fifth-generation Tensor
Cores and FP4 support, delivering up to 1000 TOPS of AI compute. GB10 includes
a high-performance Grace 20-core Arm CPU to supercharge data preprocessing
and orchestration, speeding up model tuning and real-time inferencing. The GB10
Superchip uses the NVLink™-C2C to deliver a CPU+GPU coherent memory model with
5X the bandwidth of PCIe Gen 5.

Key Features

Built on NVIDIA GB10 Grace
Blackwell Superchip
NVIDIA Blackwell GPU with
fifth-generation Tensor Core
technology
NVIDIA Grace CPU with 20-
core high-performance Arm
architecture
Up to 1000 TOPS of AI
performance using FP4
128 GB of coherent, unified
system memory
Support for up to 200 billion
parameter models
NVIDIA ConnectX™ networking
to link two systems to work
with models up to 405 billion
parameters
Up to 4 TB of NVMe storage
Compact desktop form factor
Datasheet
NVIDIA DGX Spark | Datasheet | 2
Work With Large-Parameter AI Models
With 128 GB of unified system memory and support for the FP4 data format, NVIDIA
DGX Spark can support AI models of up to 200B parameters, enabling AI developers
to prototype, fine-tune, and inference large models on their desktop. With builtin NVIDIA ConnectX network technology, two NVIDIA DGX Spark systems can be
connected to work on even larger models such as Llama 3.1 405B.
Develop Locally, Deploy Anywhere at Scale
Workflows, Models
and Tools
Development SDKs
Frameworks Services
CUDA, CUDA-X, RTX
Toolkits and Libraries
DGX OS
System Software
Ubuntu
Hardware
NGC and Bare Metal
NIM AI Blueprints AI Workbench
cuDNN cuBLAS TensorRT NCCL
NVIDIA
Kernel
Desktop RTX
GPU Drivers
Docker
Preconfigured
NVIDIA
Container Tk
CPU GPU NVDEC NVENC OFA CX
User Downloadable Preconfigured*
PyTorch TensorFlow MATLAB
DL and ML Frameworks
Riva Holoscan Metropolis Isaac
Platforms and Frameworks
*Preliminary software stack, subject to change
NVIDIA DGX Spark software stack
NVIDIA DGX Spark provides organizations and developers with a powerful,
economical experimentation ground for prototype models, freeing up valuable
compute resources in their cluster environments better suited for training
and deploying production models. Leveraging the NVIDIA AI platform software
architecture makes it possible for NVIDIA DGX Spark users to seamlessly move
their models from their desktop to DGX Cloud or any accelerated cloud or data
center infrastructure with virtually no code changes, making it easier than ever to
prototype, fine-tune, and iterate.

Technical Specifications*
Architecture NVIDIA Grace Blackwell
GPU NVIDIA Blackwell Architecture
CPU 20 core Arm, 10 Cortex-X925

• 10 Cortex-A725 Arm CUDA Cores NVIDIA Blackwell Generation Tensor Cores 5th Generation RT Cores 4th Generation Tensor Performance1 1000 AI TOPS System Memory 128 GB LPDDR5x, unified system memory Memory Interface 256-bit Memory Bandwidth 273 GB/s Storage 1 or 4 TB NVME.M2 with self-encryption USB 4x USB TypeC Ethernet 1x RJ-45 connector 10 GbE NIC ConnectX-7 Smart NIC Wi-Fi WiFi 7 Bluetooth BT 5.3 w/LE Audio-output HDMI multichannel audio output Power Consumption TBD Display Connectors 1x HDMI 2.1a NVENC | NVDEC 1x | 1x OS NVIDIA DGX™ OS System Dimensions 150 mm L x 150 mm W x 50.5 mm H System Weight 1.2 kg

• preliminary specifications, subject to change

1. Theoretical FP4 TOPS using the sparsity feature.