NVIDIA DGX Spark 4 TB: AI Supercomputer

The NVIDIA DGX Spark is a compact AI supercomputer. A generous 128GB RAM configuration enables work with large LLMs with up to 200B parameters.

The core of the NVIDIA DGX Spark is the NVIDIA® GB10 Grace Blackwell Superchip, with 1 Petaflop of AI computing power (in FP4 precision).

The NVIDIA DGX Spark platform is particularly suitable for AI developers and AI research. Thanks to its compact form factor, the NVIDIA DGX Spark fits on any desk.

An essential part of the platform is the NVIDIA AI Software Stack, which enables a quick start with the best-supported platform in the LLM industry. Thanks to software compatibility with NVIDIA's larger platforms (e.g., H200 / B200), developed LLM solutions and applications (e.g., agentic AI) can be scaled directly and easily.

Thanks to the built-in NVIDIA® ConnectX-7 network link, two NVIDIA DGX Spark systems can be connected to support even larger models with up to 405B parameters.

Delivery time
Small quantities available directly from stock in Leipzig, Germany. The delivery time for larger quantities of the NVIDIA DGX Spark 4 TB is usually 4–6 weeks from receipt of order. Please consider this in your project planning.

Technical Specifications

Scope of Delivery

• 1 × NVIDIA DGX Spark with 4 TB enterprise grade SSD

• Power Supply

• Power Cable (EU)

• Note: QSFP cable for the ConnectX CX-7 network interface, for connecting two NVIDIA DGX Spark systems, is not included - it is available separately. Compatible cables are:

• • DGX Spark Connect Cable: NJAAK-N911
  • DGX Spark Connect Cable: NJAAK-0006

Note: when ordering two NVIDIA DGX Spark 4 TB, we include a free QSFP cable during the promotion period.

Areas of Application

• Industry & Research: Development and finetuning of AI LLM models with up to 200B parameters (or 405B when using two NVIDIA DGX Sparks)
  

• Universities & Colleges: Experiments and hands-on experience for students with large LLM models

• Offices, Business Environment: Running Agentic AI, especially for data protection - e.g., for processing customer inquiries, offline email processing, RAGs with sensitive company data.

NVIDIA DGX Spark Product Variants

The NVIDIA DGX Spark platform is based on NVIDIA components at its core. Each variant features the Grace Blackwell Superchip GB10, as well as 128 GB RAM with 273 GB/s memory bandwidth. Memory bandwidth has the greatest effect on token/s throughput rates.

The operating system (NVIDIA DGX OS, based on Ubuntu) and the NVIDIA AI software stack are the same in all cases.

Different providers build a housing around it and market the device under their own brand name:

• NVIDIA directly as "NVIDIA DGX Spark" (formerly Project Digits) / DGX Spark Founders Edition
• NVIDIA DGX Spark Founder's Edition
• PNY is in this case a reseller of NVIDIA, i.e. a PNY DGX Spark is an original NVIDIA DGX Spark
• ASUS as "ASUS Ascent GX10"
• HP as "HP ZGX Nano AI Station G1n"
• Dell as "Dell Pro Max With GB10"
• Lenovo as "Lenovo ThinkStation PGX"

In our opinion, these devices are largely comparable in terms of AI performance. We recommend that customers buy the brand of their choice and pay particular attention to other factors such as service from the respective brand.

As an ASUS IoT partner, we also offer the ASUS Ascent GX 10 variant of the NVIDIA DGX Spark in our shop. ASUS's strength is in developing optimized cooling concepts, which they have also implemented with the DGX Spark in the form of the ASUS Ascent GX 10. Another advantage of the ASUS Ascent GX10 is its lower price (4 TB variant of the ASUS Ascent GX 10 available on request, also more affordable).

Software Stack / Supported Software

• NVIDIA DGX OS (Operating system, based on Ubuntu)
• NVIDIA NIM (Microservices as building blocks for AI solutions)
• NVIDIA AI Blueprints (Example applications, e.g., RAG workflows)
• PyTorch (Framework for Machine Learning and Deep Neural Networks)
• TensorFlow (Framework for Machine Learning and Deep Neural Networks)
• NVIDIA Riva (Speech recognition, translation, and speech synthesis)
• NVIDIA Holoscan (Platform for sensor data processing)
• NVIDIA Metropolis (Vision AI application platform)
• NVIDIA Isaac (Platform for developing robots, e.g., autonomous vehicles)
• CUDA (Toolkit for accelerating and parallelizing computations on GPUs)
• cuDNN (Acceleration of Deep Neural Networks on GPUs)
• cuBLAS (Linear Algebra Frameworks for accelerating AI and High Performance Computing (HPC) applications)
• TensorRT (SDK for high-performance deep learning inferencing)
• NCCL (NVIDIA Collective Communications library, for distributing computing power across multiple nodes)
• vLLM + ollama + llama.cpp - Large Language Model Inferencing Frameworks

Other Product Designations

• DGXSPARK-FOUNEDIT-EU
  
• PNY Part Number: NVDGXSPARK-PB
• PNY NVIDIA DGX Spark GB10 128GB 1 Petaflop FP4 AI Compute 128 Gb memory 4TB enterprise grade storage CX7 networking EU cable
• NVIDIA DGX Spark Founder's Edition
• 940-54242-0000
  

Safety information

• Please read the instruction manual carefully before using the product.
• Ensure that all manufacturer's assembly and installation instructions are carefully followed.
• Use the product only for its intended purpose.
• Improper use of this product may result in serious injury or property damage.
• Not suitable for children under 10 years.
• Improper use may result in injury.
• This product complies with the applicable safety requirements of the European Union.
• This product has been tested in accordance with the GPSR, which ensures that all relevant safety requirements for consumer products are met.

Traceability information

Each product has one or more of the following features:

• A CE marking indicating compliance with the safety, health and environmental protection requirements of the European Union.
• A unique serial or batch number to ensure traceability and support recalls if necessary.
• Manufacturer and importer information for customer support and safety inquiries.

Monitoring and reporting of incidents

In the unlikely event of a product issue, we have implemented procedures to:

• Handle customer complaints promptly.
• Report serious incidents via the EU Safety Gate/RAPEX system.
• Cooperate with market surveillance authorities to ensure public safety.

Contact:

• Email: support [@] pi3g.com
• Phone: 0341 / 392 858 40

This product is fully compliant with all applicable EU regulations to ensure the safety of our valued customers.