We use cookies to make your experience better. To comply with the new e-Privacy directive, we need to ask for your consent to set the cookies. Learn more.
NVIDIA DGX Spark
NVIDIA DGX Spark — Grace Blackwell AI supercomputer on your desk
NVIDIA DGX™ Spark brings the NVIDIA GB10 Grace Blackwell Superchip, 128 GB of coherent unified LPDDR5x memory, and up to 1 petaFLOP of FP4 AI performance into a compact 150 mm × 150 mm × 50.5 mm desktop appliance. Designed for AI developers and teams, it lets you prototype, fine-tune, and run large language and vision models of up to 200 billion parameters locally — with a seamless path to DGX Cloud and larger data center clusters.
Server Simply supplies DGX Spark as part of integrated AI stacks — from a single developer unit to connected pairs with high-speed NVIDIA ConnectX-7 networking and secured storage and backup options.
Petaflop-class desktop AI
Grace Blackwell GB10 combines Arm CPU and Blackwell GPU cores with fifth-generation Tensor Cores to deliver up to 1 PFLOP FP4 / ~1000 AI TOPS in a quiet desktop-friendly form factor. Ideal for LLMs, VLMs, and complex data science workflows without relying on shared cloud GPUs.
128 GB unified memory
128 GB LPDDR5x unified system memory at 273 GB/s removes the usual “VRAM wall”. Keep parameters, activations, and context in a single coherent memory space, enabling fine-tuning of models up to ~70B parameters and inference on models up to ~200B directly on the device.
Connect, pair, and scale
With 10 GbE onboard, an NVIDIA ConnectX-7 SmartNIC for high-speed clustering, Wi-Fi 7, and four USB4 Type-C ports, DGX Spark is ready for lab networks, edge deployments, or direct pairing of two systems to serve even larger models (up to ~405B parameters) as your workloads grow.
Where NVIDIA DGX Spark fits
- Local LLM and RAG development using open and proprietary models with sensitive data kept on-premises
- Fine-tuning models up to ~70B parameters without distributed training complexity
- High-throughput inference on models up to ~200B parameters for internal assistants and copilots
- Data science acceleration with RAPIDS — ETL, feature engineering, and analytics in unified GPU memory
- Robotics, vision, and edge AI prototypes using frameworks like Isaac™, Metropolis, and Holoscan
- Paired DGX Spark nodes as a mini-cluster for labs, research groups, or branch offices
- Seamless promotion of workloads to DGX Cloud or larger on-prem GPU clusters when needed
NVIDIA DGX Spark — key technical specifications
| Specification | NVIDIA DGX Spark |
|---|---|
| Architecture | NVIDIA Grace Blackwell GB10 Superchip (CPU + GPU on a single package) |
| CPU | 20-core Armv9 CPU — 10× Cortex-X925 performance cores + 10× Cortex-A725 efficiency cores |
| GPU & accelerators | NVIDIA Blackwell GPU architecture with Blackwell-generation CUDA cores, 5th-gen Tensor Cores, and 4th-gen RT Cores |
| AI performance | Up to 1 petaFLOP FP4 AI performance (≈ 1000 AI TOPS with sparsity) |
| System memory | 128 GB LPDDR5x coherent unified system memory, 256-bit interface, up to 273 GB/s bandwidth |
| Storage | 4 TB NVMe M.2 SSD with self-encryption (factory-configured); options for 1 TB or 4 TB depending on SKU |
| Networking |
|
| Wireless & I/O |
|
| Operating system & software | NVIDIA DGX™ OS with Linux, preinstalled NVIDIA AI software stack (CUDA-X libraries, containers, frameworks, NIM microservices, AI Workbench, RAPIDS, PyTorch, JAX, TensorFlow, and others). |
| Supported model sizes (guidance) |
|
| Physical characteristics | Compact desktop chassis — 150 mm L × 150 mm W × 50.5 mm H; approx. 1.2 kg |
| Power | External 240 W power supply; typical system power ~170 W, up to ~220 W under peak load. Designed for quiet operation and standard office power and cooling. |
Server Simply ships NVIDIA DGX Spark with region-appropriate power cords, validated cables for high-speed networking, and optional integration with top-of-rack switching, storage, and backup solutions on request.
DGX Spark video overview
Server Simply + DGX Spark: what you gain
Datacenter power, desktop footprint
Bring DGX-class AI performance directly into developer spaces. DGX Spark fits on a desk yet provides enough compute and memory to handle serious LLM and multimodal workloads without a dedicated server room.
Private & compliant AI
Run sensitive workloads on hardware you physically control. Combine DGX Spark with your own identity, network segmentation, and logging to meet internal security and regulatory requirements.
Fast iteration for developers
Preinstalled NVIDIA AI stack, NIM microservices, and AI Workbench let teams get from “unboxed” to “first model online” quickly, without wrestling with drivers or CUDA versions.
Ready for RAG and internal copilots
Attach DGX Spark to your existing storage and search systems and build retrieval-augmented generation over internal documents, codebases, and knowledge bases — fully under your control.
Scale out when needed
Use DGX Spark for day-to-day development and on-prem inference, then hand off final training or large-scale experiments to DGX Cloud or GPU clusters without changing tooling.
Single partner for the stack
Server Simply can supply not only DGX Spark itself, but also compatible Supermicro or NVIDIA-based infrastructure for storage, backup, networking, and future cluster growth.
Think of DGX Spark as your on-desk AI factory: develop, fine-tune, and serve models locally — then scale to racks or cloud when it’s time to go from prototype to production.
Start your DGX Spark journey with Server Simply
NVIDIA DGX Spark makes petaflop-class AI accessible at the desk. Server Simply helps you turn it into a complete solution — with sizing, networking, storage, and lifecycle support tailored to your AI roadmap. Tell us about your workloads, and we’ll propose a DGX Spark–based design that fits your budget, compliance posture, and growth plans.
