NVIDIA launches DGX Spark: a pocket-sized AI supercomputer for developers
NVIDIA has begun selling the DGX Spark mini PC for $3,999. At just 2.6 pounds, the Spark packs NVIDIA’s GB10 super system-on-chip — combining a 20‑core ARM CPU with a Blackwell GPU — plus 128GB of unified LPDDR5x memory and 4TB of NVMe storage. NVIDIA positions the device as a compact development workstation for researchers and data scientists rather than a consumer gaming PC.
Key specs at a glance:
- GB10 SoC: 20‑core ARM CPU + Blackwell GPU
- 128GB LPDDR5x unified memory (shared between CPU & GPU)
- 4TB NVMe storage
- Connectivity: Wi‑Fi 7, four USB‑C ports, HDMI
- Runs NVIDIA DGX OS (a tuned Ubuntu variant with NVIDIA’s AI stack)
The DGX Spark runs DGX OS, giving developers access to NVIDIA’s models, libraries and microservices for tasks like fine‑tuning image generators or building chatbots. Early recipients include companies and research teams such as Anaconda, Google, Hugging Face, Meta and Microsoft — and NVIDIA CEO Jensen Huang even hand‑delivered a unit to Elon Musk at SpaceX’s Starbase.
The Spark is part of a broader move toward compact, AI‑focused developer machines. At Computex, multiple vendors (Dell, HP, Lenovo, ASUS and others) demoed similar mini PCs built around the same GB10 platform; Acer’s Veriton GN100 was one example. NVIDIA is also developing a larger DGX Station model using the GB300 Grace Blackwell Ultra superchip, which promises far higher performance (20 petaflops and 784GB unified memory), though pricing for that model has not yet been announced.
At $3,999 the DGX Spark is not cheap, but it offers a high density of AI‑ready components in a very small footprint — making it a viable option for labs, startups and teams that need dedicated local hardware for model development and testing.
For more details, see the original coverage on Engadget or visit NVIDIA.
Discussion: Would a compact AI workstation like the DGX Spark change how you prototype or train models locally? Share your thoughts below.