DGXSPARK-FOUNEDIT

Overview

PNY NVIDIA DGX Spark Desktop AI Supercomputer
The PNY NVIDIA DGX Spark brings powerful AI computing into a compact desktop system, giving developers, researchers and data scientists the ability to prototype, fine-tune and run large AI models locally. Built around the NVIDIA Grace Blackwell architecture, it delivers up to 1 petaFLOP of FP4 AI performance while combining a 20-core Arm CPU, Blackwell GPU technology and 128 GB of unified LPDDR5x memory in a remarkably small chassis.

With support for AI models of up to 200 billion parameters on a single system, the DGX Spark is designed for demanding generative AI, machine learning, inference and development workloads. Two units can also be connected using NVIDIA ConnectX networking to support models of up to 405 billion parameters, providing a practical way to expand local AI computing without moving immediately to a traditional server environment.

Description

The PNY NVIDIA DGX Spark is a desktop AI supercomputer designed to give AI developers access to high-performance computing without the size and infrastructure demands of a conventional data centre system. Its NVIDIA GB10 Superchip combines a 20-core Arm CPU with a Blackwell-generation GPU, 5th-generation Tensor Cores and 4th-generation RT Cores for demanding AI workloads.

The 128 GB unified LPDDR5x memory is shared coherently between the CPU and GPU, helping reduce memory limitations when working with large language models and other data-intensive applications. With 273 GB/s of memory bandwidth, the system can handle large datasets, model checkpoints and local AI workloads efficiently.

A 4 TB self-encrypting NVMe M.2 drive provides substantial high-speed storage for AI models, datasets and development environments. The system also includes the NVIDIA AI software stack through DGX OS, providing a ready-to-use environment for AI development, model training and inference.

Features

Desktop AI Computing with Grace Blackwell
The DGX Spark uses NVIDIA Grace Blackwell architecture to provide data-centre-class AI computing in a compact desktop form factor. Its 1 petaFLOP FP4 performance is aimed at workloads such as large language model development, generative AI, transformer-based applications and inference.

The integrated 20-core Arm processor combines 10 Cortex-X925 performance cores with 10 Cortex-A725 cores, working alongside the Blackwell GPU architecture to provide a balanced platform for CPU and GPU accelerated workloads.

Key processing capabilities include:

• NVIDIA Grace Blackwell architecture
• Blackwell-generation GPU
• 20-core Arm CPU
• 10 Cortex-X925 cores and 10 Cortex-A725 cores
• 5th-generation Tensor Cores
• 4th-generation RT Cores
• Up to 1 petaFLOP FP4 Tensor performance

128 GB Unified Memory for Large AI Models
One of the main advantages of the DGX Spark is its 128 GB of unified LPDDR5x system memory. Rather than treating CPU and GPU memory as completely separate pools, the architecture allows them to share the available memory coherently.

This large memory capacity is particularly useful when developing and running large AI models locally. A single DGX Spark can support AI models of up to 200 billion parameters, while connecting two systems can extend the capability to models of up to 405 billion parameters.

Memory details include:

• 128 GB LPDDR5x unified memory
• 256-bit memory interface
• 273 GB/s memory bandwidth
• Support for AI models up to 200 billion parameters on one system
• Support for AI models up to 405 billion parameters when two systems are connected

Fast Storage for AI Development
Large AI models and datasets can quickly consume storage capacity, so the DGX Spark includes a 4 TB NVMe M.2 drive. The self-encrypting storage provides additional protection for locally stored models, datasets and other valuable development data.

The combination of high-capacity NVMe storage and unified system memory makes the DGX Spark well suited to local AI development where large datasets and model checkpoints need to remain readily accessible.

High-Speed Networking and Connectivity
Despite its compact size, the DGX Spark provides connectivity suited to demanding AI environments. A 10 GbE RJ-45 connection provides fast wired networking, while the NVIDIA ConnectX-7 Smart NIC can support high-speed networking for multi-system configurations.

Wireless connectivity is also included, with Wi-Fi 7 and Bluetooth 5.3 providing flexible options for peripherals and network access.

Connectivity includes:

• 1x 10 GbE RJ-45 Ethernet connection
• NVIDIA ConnectX-7 Smart NIC
• Up to 100 GbE networking capability through ConnectX-7
• Wi-Fi 7
• Bluetooth 5.3
• 4x USB Type-C ports
• 1x HDMI 2.1a display output
• HDMI multichannel audio output
• Support for displays up to 8K at 120 Hz with HDR

Compact Desktop Design
The DGX Spark is designed to provide substantial AI computing power without requiring a traditional rack-mounted server setup. Measuring approximately 15 x 15 x 5 cm and weighing around 1.2 kg, it can fit easily into a desktop development environment.

The compact champagne-gold enclosure gives the system a distinctive appearance while keeping the physical footprint small. This makes it practical for AI developers, researchers and professional users who need local computing power but have limited workspace.

Built for AI Development and Research
The DGX Spark is aimed at users who need to experiment with increasingly large AI models without depending entirely on remote cloud infrastructure. Developers can use it for prototyping, fine-tuning, inference, retrieval-augmented generation and multimodal AI workloads.

The preinstalled DGX OS and NVIDIA AI software environment are designed to provide the software foundation needed for AI development, including frameworks, libraries and pretrained models.

Typical applications include:

• Large language model development
• Generative AI development
• AI model prototyping and fine-tuning
• Local AI inference
• Transformer-based workloads
• Retrieval-augmented generation
• Multimodal AI applications
• Machine learning research
• AI development and experimentation
• High-performance desktop AI computing

Key Technical Details

• Architecture: NVIDIA Grace Blackwell
• GPU: Blackwell architecture
• CPU: 20-core Arm, 10 Cortex-X925 + 10 Cortex-A725
• Tensor Cores: 5th generation
• RT Cores: 4th generation
• Tensor Performance: Up to 1 PFLOP FP4
• System Memory: 128 GB LPDDR5x unified memory
• Memory Interface: 256-bit
• Memory Bandwidth: 273 GB/s
• Operating System: DGX OS
• Storage: 4 TB NVMe M.2 with self-encryption
• USB: 4x USB Type-C
• Ethernet: 1x 10 GbE RJ-45
• Network Interface: NVIDIA ConnectX-7 Smart NIC
• Wireless: Wi-Fi 7
• Bluetooth: Bluetooth 5.3
• Display: 1x HDMI 2.1a
• Audio: HDMI multichannel audio output
• NVENC: 1x
• NVDEC: 1x
• Dimensions: 15 x 15 x 5 cm
• Weight: 1.2 kg
• Power Consumption: TBD

Part Number and Product Identification
For ordering and product identification, the PNY NVIDIA DGX Spark is supplied under part number / SKU DGXSPARK-FOUNEDIT. This identifier can be used when checking availability, compatibility and purchasing requirements.

What's Included
The package includes the essential hardware and documentation required to get the DGX Spark started:

• NVIDIA DGX Spark - Gold/Champagne x1
• Power Supply x1
• Quick Start Card x1
• Support Card x1

A Compact AI Supercomputer for Local Workloads
The PNY NVIDIA DGX Spark is a specialised desktop AI system for users who need more than a conventional workstation can provide. Its combination of Grace Blackwell architecture, 128 GB unified memory, 1 petaFLOP FP4 performance, 4 TB NVMe storage and high-speed networking makes it particularly suited to large-scale AI development and experimentation.

For developers, researchers and organisations looking to bring more AI workloads onto the desktop, the DGX Spark provides a compact platform that can handle large models locally while also offering a path towards multi-system AI computing.

Specifications