What is NVIDIA DGX Spark?

NVIDIA DGX Spark is a pint-sized “AI supercomputer” for your desk. Built around NVIDIA’s GB10 Grace Blackwell Superchip, it’s designed to let developers and researchers prototype, fine-tune, and run inference on large AI models locally without booking time on a data center cluster. It offers up to 1 petaFLOP (FP4) of AI performance and 128GB of unified memory in a compact form factor, with NVIDIA’s AI software stack preinstalled.

• GB10 Grace Blackwell Superchip (Grace CPU + Blackwell GPU in one package)
• 128GB LPDDR5x unified system memory (CPU and GPU share it coherently via NVLinkC2C)
• NVIDIA ConnectX networking (10GbE onboard; ConnectX7 SmartNIC)
• Up to 4TB NVMe storage
• Tiny footprint: roughly 150 x 150 x 50.5 mm; about 1.2 kg
• DGX OS + NVIDIA AI software stack out of the box

DGX Spark is built for the “make it work on my desk” phase of AI:

• Prototyping: Build and validate models and AI-augmented apps locally, then hand them off to bigger infrastructure if needed.
• Fine-Tuning: Tweak models with up to ~70B parameters directly on the box.
• Inference: Run state-of-the-art models with up to ~200B parameters for testing and validation. Link two DGX Spark units via ConnectX to push into ~405B parameter territory.
• Data Science: Accelerate end-to-end pipelines with NVIDIA RAPIDS (and even boost Apache Spark with the RAPIDS Accelerator).
• Edge and Robotics Development: Experiment with frameworks like Isaac, Metropolis, and Holoscan on a deskside system.

• Unified Memory vs. Separate VRAM: DGX Spark’s 128GB is coherent system memory shared between the CPU and GPU, making it ideal for large context windows and efficient data movement. Conventional PCs split RAM and GPU VRAM.
• AI-First Silicon: The GB10’s fifth-generation Tensor Cores and FP4 support are designed specifically for modern LLMs and AI agents. This is not a frames-per-second machine.
• Stack Included: DGX OS and NVIDIA’s AI platform come preinstalled, so you’re much closer to “open notebook, run model” than “install drivers, hunt for containers.”

No, DGX Spark is a hardware system, while Apache Spark is a distributed data processing framework. The nice part is that if you do use Apache Spark, NVIDIA’s RAPIDS Accelerator for Apache Spark can offload parts of your pipelines to the GPU, and DGX Spark supports that stack.

NVIDIA’s official product page for DGX Spark focuses on specs and signups, while availability runs through NVIDIA and partner OEMs. NVIDIA announced that Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, and MSI will offer DGX Spark systems, with availability starting in July (regional rollout varies).

As for pricing, reports suggest configurations starting at around $3,999, though final prices depend on the OEM and storage options. Some retail pages still show “coming soon,” so check partner listings for current pricing and stock.

If DGX Spark is your deskside development box, DGX Station is the desktop AI powerhouse. DGX Station (GB300 Ultra) targets the most demanding training and fine-tuning jobs, delivering up to ~20 petaFLOPs (FP4) and hundreds of gigabytes of unified memory. It is much larger and designed for teams or shared lab environments.

Yes if you’re an AI developer, data scientist, or researcher who constantly iterates on LLMs, agents, or multimodal models and needs fast local turnaround, private data handling, and a software stack that maps cleanly to the data center or cloud.

Maybe not, if your needs are limited to GPU rendering or gaming, or if you already have steady access to cluster or HPC time. In that case, a traditional workstation or cloud credits could be more cost effective.

• AI performance: up to 1 PFLOP (FP4)
• Memory: 128GB LPDDR5x unified (273 GB/s)
• Storage: 1TB or 4TB NVMe (self-encrypting)
• Networking: 10GbE, ConnectX-7 SmartNIC, Wi-Fi 7, BlueTooth 5.3
• I/O: 4x USB-C, 1x HDMI 2.1a, NVENC/NVDEC (1/1)
• Size and Weight: 150 x 150 x 50.5 mm; ~1.2 kg
• OS: NVIDIA DGX OS