NVIDIA DGX Spark Personal AI Supercomputer

Description

The NVIDIA DGX Spark is a purpose-built desktop AI computing platform that brings the NVIDIA Grace Blackwell architecture into a compact, low-power system.

Unlike a conventional workstation that combines a separate CPU, discrete graphics card, and system memory, DGX Spark uses the GB10 Grace Blackwell Superchip, integrating CPU and GPU resources with a coherent 128GB unified memory architecture. This enables the CPU and GPU to access the same large memory pool, which is particularly useful for AI models that exceed the memory capacity of conventional consumer GPUs.

GB10 Grace Blackwell Superchip

At the center of DGX Spark is NVIDIA’s GB10 Grace Blackwell Superchip.

The CPU portion contains 20 Arm cores, consisting of:

• 10 × Cortex-X925 high-performance cores
• 10 × Cortex-A725 efficiency cores

The GPU uses the Blackwell architecture, with:

• 6,144 CUDA cores
• Fifth-generation Tensor Cores
• Fourth-generation RT Cores
• FP4 acceleration
• Up to 1 PFLOP of AI compute with sparsity

This combination provides substantial AI processing capability while maintaining a compact desktop form factor.

128GB Unified Memory

One of DGX Spark’s defining features is its 128GB LPDDR5x coherent unified memory.

Rather than having separate CPU system RAM and GPU VRAM, the CPU and GPU share the same memory pool. The memory subsystem provides:

• 128GB LPDDR5x
• 256-bit interface
• Up to 273GB/s memory bandwidth
• 16 memory channels

This architecture allows large AI models to be loaded directly into shared CPU/GPU memory and reduces the need for data movement between separate CPU and GPU memory pools.

AI Performance

DGX Spark delivers up to 1 PFLOP of FP4 AI performance using sparsity, equivalent to up to 1,000 TOPS of inference performance under NVIDIA’s stated theoretical configuration.

The platform is optimized for:

• LLM inference
• Generative AI
• Model fine-tuning
• Computer vision
• Multimodal AI
• AI agents
• Data science
• Robotics development
• Edge AI

NVIDIA positions DGX Spark for local inference of models up to 200 billion parameters and local fine-tuning workloads of up to approximately 70 billion parameters, subject to workload and software requirements.

High-Speed Networking

DGX Spark integrates a ConnectX-7 Smart NIC operating at up to 200Gb/s, providing high-bandwidth connectivity for distributed AI workloads and communication with other accelerated systems.

The system also includes:

• 10GbE RJ-45 networking
• Wi-Fi 7
• Bluetooth 5.4
• Two QSFP network connections associated with the ConnectX-7 subsystem

This allows DGX Spark to operate as an individual AI development system or participate in larger networked AI environments.

NVLink-C2C

The GB10 architecture uses NVLink-C2C to provide high-bandwidth communication between CPU and GPU resources.

NVIDIA describes the technology as providing substantially greater CPU-GPU bandwidth than conventional PCIe-based discrete GPU configurations, enabling the unified memory architecture to support data-intensive AI workloads efficiently.

Storage

The current NVIDIA DGX Spark specification includes a 4TB NVMe M.2 SSD with self-encryption.

The high-capacity NVMe storage provides space for:

• AI models
• Datasets
• Development environments
• Containers
• CUDA libraries
• AI frameworks
• Project files

NVIDIA documentation also references 1TB and 4TB storage configurations for DGX Spark/GB10-based systems; the current NVIDIA Founders Edition specification lists 4TB.

NVIDIA DGX Software Ecosystem

DGX Spark is delivered as a complete AI computing platform rather than simply a hardware appliance.

It runs NVIDIA DGX OS and supports NVIDIA’s AI software ecosystem, including CUDA and optimized AI frameworks and tools.

The platform is designed to work with:

• CUDA
• PyTorch
• TensorRT
• TensorRT-LLM
• NVIDIA NIM
• NVIDIA AI Enterprise
• NVIDIA AI frameworks
• Containerized AI workloads
• Generative AI models

The current DGX Spark software stack includes NVIDIA DGX OS, CUDA, GPU drivers, and the associated NVIDIA AI development environment.

3. Complete Technical Specification

The specifications above reflect NVIDIA’s current DGX Spark documentation.

4. Applications

Artificial Intelligence & Machine Learning

• Large language model inference
• Generative AI development
• AI agent development
• Local model experimentation
• Model fine-tuning
• Multimodal AI
• Computer vision
• Speech and language AI

Data Science

• Large-scale data analysis
• Machine learning development
• Python-based AI workflows
• Dataset processing
• Local experimentation
• GPU-accelerated analytics

AI Research & Development

• Prototype development
• Model validation
• Algorithm development
• AI benchmarking
• Research environments
• Edge AI development

Robotics & Edge Computing

DGX Spark can be used to develop applications based on NVIDIA AI platforms such as:

• NVIDIA Isaac
• NVIDIA Metropolis
• NVIDIA Holoscan
• Computer vision
• Robotics
• Autonomous systems
• Smart-city applications

NVIDIA specifically identifies DGX Spark as a platform for developing edge and physical-AI applications.

Professional AI Development

• AI software development
• LLM application development
• RAG systems
• AI agents
• Generative image/video development
• Local inference servers
• AI education and training

5. Compatibility

NVIDIA DGX Spark is designed as a complete integrated AI computer, rather than a conventional workstation requiring a separate CPU, GPU, motherboard, and memory configuration.

It is compatible with NVIDIA’s AI software ecosystem, including:

• NVIDIA CUDA
• PyTorch
• TensorRT
• TensorRT-LLM
• NVIDIA NIM
• NVIDIA AI Enterprise
• NVIDIA DGX OS
• NVIDIA AI development frameworks
• Containerized AI applications

The system can also be connected to:

• 10GbE networks
• High-speed ConnectX-7 networks
• Wi-Fi 7 networks
• External USB-C devices
• HDMI displays
• DisplayPort displays through USB-C DP Alt Mode
• Networked AI infrastructure
• NVIDIA cloud and data-center environments

NVIDIA describes DGX Spark as a platform that allows developers to prototype and develop locally before moving workloads to DGX Cloud or accelerated data-center infrastructure.

6. Key Benefits

• 1 PFLOP FP4 AI performance
• 6,144 CUDA cores
• NVIDIA Blackwell architecture
• 20-core Arm CPU
• 128GB coherent unified memory
• High-bandwidth 273GB/s memory subsystem
• Integrated GB10 Grace Blackwell Superchip
• NVLink-C2C CPU-GPU interconnect
• Integrated ConnectX-7 200Gb/s networking
• Built-in 10GbE
• Wi-Fi 7 and Bluetooth 5.4
• 4TB self-encrypting NVMe storage
• Compact 150 × 150 × 50.5mm desktop form factor
• Only approximately 1.2kg
• 240W external power supply
• NVIDIA DGX OS and optimized AI software stack
• Designed for models up to approximately 200B parameters for inference
• Ideal for local AI development without relying exclusively on cloud GPUs
• Supports AI, data science, robotics, computer vision and generative AI workflows

7. SEO Title

NVIDIA DGX Spark AI Supercomputer – GB10 Grace Blackwell, 128GB Unified Memory, 1 PFLOP