MiaAI-Lab/sparkDash

What it solves

sparkDash is a real-time monitoring dashboard designed for NVIDIA DGX Spark (GB10) machines and other Linux hosts with NVIDIA GPUs. It solves the problem of managing multiple AI hardware units in a single interface, providing visibility into GPU, CPU, memory, and network performance, as well as the health and performance of local LLM servers.

How it works

The system uses a Node.js/Express backend that manages a registry of "Sparks" (monitored units). It collects metrics via sysfs/proc and nvidia-smi for local hosts, and via SSH for remote units. A WebSocket stream pushes these metrics to a React-based frontend. It includes specialized probes for LLM backends (like vLLM, llama.cpp, and SGLang) and optional integrations for ComfyUI, Hermes Agent, and Tailscale to monitor service health and connectivity.

Who it’s for

It is intended for developers and researchers operating NVIDIA DGX Spark units or GPU-enabled Linux workstations who need a centralized way to monitor hardware utilization, benchmark LLM inference performance, and manage remote GPU hosts.

Highlights

• Multi-unit Monitoring: Track multiple DGX Spark or standard NVIDIA GPU hosts in one browser window with a shared overview.
• LLM Performance Probing: Auto-detects various LLM backends to report live decode/prefill tokens per second and KV cache usage.
• Inference Benchmarking: Built-in tools for multi-concurrency decode and context-size sweep prefill benchmarks (1k to 300k tokens).
• Service Integrations: Optional monitoring for ComfyUI (queue/job status), Hermes Agent (update checks), and Tailscale (off-tailnet detection).
• Hardware Detail: Specifically tracks unified memory for GB10 units and separates RAM from VRAM for discrete GPU hosts.
• Remote Management: Supports SSH-based metric collection, Wake-on-LAN, and graceful shutdown commands.