NVIDIA DGX Spark™ – Personal AI Desktop Supercomputer – Desktop GB10 Grace Blackwell Chip

Bring supercomputer-level performance directly to your desk with a compact, energy-efficient design that enables enterprise-scale AI and high-performance computing exactly where you need it. Powered by the advanced Grace Blackwell architecture, it delivers high AI performance, accelerating local AI model fine-tuning, inference, and advanced analytics.

The NVIDIA DGX Spark brings data-centre-class AI computing into a chassis the size of a hardback book. Designed for machine learning engineers, researchers, and small AI teams, it enables local prototyping, fine-tuning, and running large AI models without relying on shared clusters or costly cloud GPU instances. Instead of spinning up an A100 or H100 instance for every experiment, DGX Spark provides dedicated desktop hardware — eliminating idle-GPU billing, keeping datasets on-site, and removing delays from shared infrastructure.

Key Features

• NVIDIA GB10 Grace Blackwell Superchip — a 20-core Arm CPU (10x Cortex-X925 performance cores + 10x Cortex-A725 efficiency cores) paired with a Blackwell GPU featuring 5th-generation Tensor Cores and 4th-generation RT Cores via NVLink-C2C. Delivers up to 1 petaFLOP of FP4 AI compute with sparsity, bringing NVIDIA's data-centre silicon architecture to the desktop.

• 128GB Unified LPDDR5x Memory at 273GB/s — CPU and GPU share a coherent 128GB memory pool over a 256-bit interface, eliminating PCIe data transfers between system RAM and VRAM. Enables models that exceed the memory capacity of 24GB or 48GB discrete workstation GPUs.

• Run and Fine-Tune Models Up to ~200B Parameters — supports local inference, LoRA/QLoRA fine-tuning, and RAG development with models in the 70B–200B parameter class at FP4 precision, including fully offline or air-gapped operation.

• 4TB Self-Encrypting NVMe SSD — ample local storage for model checkpoints, quantized variants, container images, and training datasets without frequent external transfers.

• ConnectX-7 Smart NIC (200Gbps) — connect two DGX Spark units using a certified QSFP112 interconnect cable to pool 256GB of unified memory and run inference on models up to approximately 405B parameters.

• Full NVIDIA DGX OS + AI Stack, Pre-Installed — includes CUDA, cuDNN, NCCL, PyTorch and TensorFlow containers, NVIDIA AI Workbench, and frameworks such as Isaac, Metropolis, and Holoscan. Workloads are container-portable to other Blackwell systems or DGX Cloud without code changes.

• Desktop Footprint, Data-Centre Silicon — compact 150 x 150 x 50.5mm design weighing 1.2kg, powered by a single 240W supply. Includes 1x HDMI 2.1a and up to 3x DisplayPort 2.1 connections via USB-C for multi-monitor visualisation.

• Enterprise-Grade I/O — 4x USB Type-C, 10 Gigabit Ethernet (RJ-45), Wi-Fi 7, and Bluetooth 5.4 provide professional workstation connectivity.

Built for Local AI Development, Not Cloud-Dependent Demos

DGX Spark bridges the gap between laptops that cannot handle large AI models and expensive cloud GPU infrastructure. Fine-tuning a 70B open-weight model, developing RAG pipelines with private datasets, or benchmarking quantization strategies often requires dozens of short experiments — workflows that can quickly increase cloud computing costs. With 128GB of unified memory shared between CPU and GPU, DGX Spark handles models that exceed the VRAM capacity of many workstation and data-centre GPUs, without complex multi-GPU sharding.

Its unified memory architecture is a major advantage. Traditional discrete-GPU workstations require model weights, KV cache, and activations to fit within dedicated VRAM. Exceeding that capacity means offloading to slower system RAM or failing to load the model. DGX Spark instead provides direct GPU access to the shared 128GB LPDDR5x pool at 273GB/s, allowing 70B-parameter models in FP4/INT4 quantization or 30B models at higher precision with extended context lengths to run without complicated memory-management workarounds.

The advertised 1 petaFLOP represents peak FP4 performance with sparsity; actual dense-precision throughput will be lower depending on the workload. DGX Spark is not intended for training foundation models from scratch, as its 273GB/s memory bandwidth and single-GPU-equivalent compute remain better suited to fine-tuning and inference than large-scale pre-training. Its strengths lie in quantization, inference serving, agentic AI development, computer vision, and multimodal prototyping — enabling everyday ML workflows on dedicated local hardware.

Running NVIDIA DGX OS, an Ubuntu Linux Pro-based distribution with CUDA and container tools preinstalled, DGX Spark offers a software environment similar to DGX clusters and DGX Cloud. NVIDIA's enterprise-maintained ecosystem helps reduce compatibility concerns commonly associated with Arm-based workstations.

Two configurations are available: a Single Unit with 128GB of unified memory, or a Dual Unit + Interconnect Cable bundle featuring two DGX Spark systems and an NVIDIA-certified ConnectX-7 QSFP112 cable. The dual configuration pools 256GB of unified memory for inference on models up to approximately 405B parameters. Select the configuration based on the largest models you expect to run over the system's lifetime.

Specifications

• Superchip: NVIDIA GB10 Grace Blackwell
• CPU: 20-core Arm (10x Cortex-X925 + 10x Cortex-A725)
• GPU: NVIDIA Blackwell architecture, 5th Gen Tensor Cores, 4th Gen RT Cores
• CPU-GPU Interconnect: NVLink-C2C
• AI Performance: Up to 1 petaFLOP (FP4, with sparsity)
• System Memory: 128GB LPDDR5x, coherent unified memory
• Memory Interface / Bandwidth: 256-bit / up to 273GB/s
• Storage: 4TB NVMe M.2, self-encrypting
• Networking: NVIDIA ConnectX-7 Smart NIC (200Gbps), 10 Gigabit Ethernet (RJ-45), Wi-Fi 7, Bluetooth 5.4
• USB: 4x USB Type-C
• Display Outputs: 1x HDMI 2.1a, up to 3x DisplayPort 2.1 (USB-C DP Alt Mode)
• Video Encode/Decode: 1x NVENC / 1x NVDEC
• Operating System: NVIDIA DGX OS with full NVIDIA AI software stack (CUDA, cuDNN, NCCL, container runtime)
• Max Model Size: ~200B parameters (single unit); ~405B parameters (dual-unit clustered, 256GB pooled memory)
• Power Supply: 240W (GB10 TDP 140W)
• Dimensions: 150 x 150 x 50.5m