Nvidia DGX Spark Arrives for AI Developers (nvidianews.nvidia.com)

NVIDIA began shipping DGX Spark, a compact “desktop” AI supercomputer designed to put peta-scale compute on developers’ desks. The system packs up to 1 petaflop of AI performance with 128 GB of unified CPU–GPU coherent memory, powered by the NVIDIA GB10 Grace Blackwell Superchip, NVIDIA ConnectX-7 200 Gb/s networking and NVLink‑C2C (claimed as 5× PCIe Gen5 bandwidth). NVIDIA says DGX Spark can run inference for models up to ~200 billion parameters and fine-tune models up to ~70 billion parameters locally, and it ships with the full NVIDIA AI stack (CUDA libraries, NIM microservices, preinstalled models and ecosystem tools) so teams can start building agentic and physical AI workflows out of the box. For the AI/ML community this lowers the barrier to experimentation with large models and privacy-sensitive workloads by moving high-memory, low-latency development from the cloud into labs and offices. Partners including Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, MSI and major software/platform vendors are validating tooling and models for DGX Spark, which NVIDIA begins selling Oct. 15. The device promises faster iteration cycles for prototyping, local agent/robotics development, and on-prem model customization (examples cited: image-generation fine-tuning, vision-language agents, chatbots), potentially accelerating research while reducing cloud costs and data exposure.