Nvidia DGX Spark: great hardware, early days for the ecosystem (simonwillison.net)

NVIDIA’s new DGX Spark is a compact “AI supercomputer” desktop now shipping at about $4,000. The preview unit packs ARM64 silicon (20 cores: 10x Cortex‑X925 + 10x Cortex‑A725), roughly 128 GB of unified system/GPU-accessible RAM (reportedly ~119–128 GB available), a 4 TB NVMe SSD, and an NVIDIA GB10 Blackwell GPU (sm_121 compute capability, ~119.7 GB GPU memory, 48 SMs). It’s aimed squarely at researchers for both training and inference, bringing high‑end Blackwell power into a small, affordable form factor. The catch is software maturity: CUDA on ARM plus an ARM host breaks many x86 assumptions in the ML stack. The reviewer managed to get a PyTorch 2.7 CUDA wheel working (2.8 less clear), and found NVIDIA’s CUDA containers (e.g., nvcr.io/nvidia/cuda:13.0.1-devel-ubuntu24.04) essential. Early community and vendor support is encouraging — Ollama works out of the box, llama.cpp has published strong Spark benchmarks (~3,600 tokens/s read & ~59 tokens/s generate for MXFP4 GPT‑OSS 20B; ~817 read & ~18 gen for GLM‑4.5‑Air), LM Studio and an official vLLM NGC container are available, and NVIDIA recently published much better Spark docs. Remote access via Tailscale and tooling like Claude Code eased setup. Bottom line: the hardware is impressive, but the ecosystem is in early, rapidly improving stages — worthwhile for enthusiasts and early adopters, while mainstream readiness will depend on continued tooling and wheel availability.