Skip to main content
Back

Mia AI Lab's sparkDash Watches Your Entire NVIDIA DGX Spark Fleet at Once

sparkDash is an open-source React and Express dashboard that streams live GPU, memory, and LLM metrics from any number of NVIDIA DGX Spark machines in one browser tab.

Repo, Full Access
  • sparkDash is a real-time web dashboard for monitoring multiple NVIDIA DGX Spark GB10 units.
  • Streams GPU, CPU, unified memory, storage, network, and local LLM metrics over WebSocket.
  • Auto-detects llama.cpp, vLLM, and sglang backends, reporting live tokens per second.
  • Add, edit, reorder, or remove Sparks from the UI with no process restart required.
  • Ships as privileged ARM64 Docker container; SSH passwords encrypted with AES-256-GCM.
  • MIT licensed; API is unauthenticated so run only on a trusted LAN.

sparkDash consolidates GPU, CPU, memory, network, storage, and LLM throughput data from multiple NVIDIA DGX Spark systems. Mia AI Lab’s v2.0 release replaces per-machine terminal checks with a real-time web interface where operators can add, rename, reorder, and remove monitored systems without restarting the service.

The MIT-licensed project combines a React 19 single-page application with an Express 5 server. It runs as a privileged ARM64 Docker container and mounts the host’s /proc, /sys, and nvidia-smi resources to collect hardware metrics. The v2.0 code is available on the main branch.

JSON becomes the fleet map

Each DGX Spark is represented by a record in config/sparks.json. The interface updates that file, and the server adjusts its active monitors while running. Adding another machine requires its connection details rather than application changes or a separate collector implementation.

Local and remote machines use the same SparkMonitor, SystemCollector, and LlmProbe components. Local collection reads host data through mounted files and nvidia-smi, sometimes using nsenter to execute inside the host’s namespaces. Remote collection sends the equivalent commands through a shared sshExec() helper, using an SSH agent or sshpass.

Both collection paths produce the same data model, which keeps dashboard behavior consistent as systems are added or reordered.