Nvidia's DGX Spark: Mini AI Supercomputer Overview and Review (robert-mcdermott.medium.com)

NVIDIA’s DGX Spark (originally announced as Project DIGITS) is a compact “mini” AI supercomputer built around the GB10 Grace Blackwell superchip and aimed at researchers, students and developers who need CUDA-capable, large-memory hardware on a desktop form factor. The Founder’s Edition ships with 20 Arm cores (10x Cortex‑X925 at 4.0 GHz + 10x Cortex‑A725 at 2.8 GHz), a GB10 accelerator, 128 GB of unified LPDDR5, a 4 TB SSD, and optional networking up to 200 Gbps RDMA. It boots as a desktop (Ubuntu 24.04 + DGX OS) or headless server, supports remote workflows via Tailscale and NVIDIA Sync (JupyterLab, VS Code, remote runtimes), and includes DGX Spark “playbooks” and training vouchers to get users up to speed. Founder’s Edition retail pricing rose from $2,999 to $3,999, with vendor variants following. Hands‑on benchmarks (llama.cpp with GPT‑OSS models) validate NVIDIA’s numbers: for GPT‑OSS‑20B the system achieves ~3,685 prefill tokens/s and ~85 response tokens/s; for GPT‑OSS‑120B ~1,821 prefill tokens/s and ~50 response tokens/s. Key tradeoffs: DGX Spark offers ~1,000 TOPS of compute and 128 GB of GPU‑accessible unified memory—unique at this price—but its LPDDR5 memory bandwidth (273 GB/s) is lower than Apple’s M4 Max (546 GB/s), which can affect inference latency. The machine’s real strength is enabling fine‑tuning and large‑model workflows with full CUDA support and large memory capacity—something small GPUs (e.g., RTX 5090’s 32 GB) can’t match—making it a compelling, portable option for prototyping and hands‑on model development.