Hello everyone,
I wanted to share a small project I built for monitoring the NVIDIA DGX Spark. I’ll keep updating the project as I test more things and add improvements, so feedback and suggestions are always welcome.
The problem
The DGX Spark uses a unified-memory architecture. The GPU has no discrete VRAM: nvidia-smi reports memory as Not Supported / N/A, and every CUDA allocation lives in the host’s system RAM. Standard GPU exporters have nothing useful to scrape for memory, and a plain node exporter doesn’t understand GPU processes.
What I built using Kimi K3
A single Docker Compose stack. One small handwritten python collector runs with the NVIDIA container runtime, reads /proc/meminfo plus nvidia-smi. Prometheus scrapes it and Grafana visualizes everything on an auto-provisioned dashboard.
Quick start
git clone https://github.com/shahizat/dgx-spark-dashboard
cd dgx-spark-dashboard
docker compose up -d
docker compose ps
Then open:
- Grafana: http://localhost:3000 (admin / admin)
- Prometheus: http://localhost:9090
Remote access (optional): show the dashboard from anywhere
Grafana listens on localhost, which is fine for the lab. If you want to check the dashboard from a phone or laptop while away, there are two clean options: Cloudflare Tunnel or Tailscale Funnel.
Option A: Cloudflare Tunnel
Install the cloudflared client, then the quickest path is a TryCloudflare temporary tunnel:
cloudflared tunnel --url http://localhost:3000
It prints a random https://<random>.trycloudflare.com URL you can open anywhere. The URL changes on every restart, so for a stable hostname use a named tunnel on a domain you own:
Option B: Tailscale Funnel
If you run Tailscale, Funnel exposes a local port to the internet and provisions an HTTPS certificate + MagicDNS hostname for you. On the DGX Spark, with Tailscale running and logged in:
tailscale funnel 3000
Tailscale prints a public https://<machine-name>.<tailnet>.ts.net URL, and routes traffic back over the encrypted Tailscale network.
I’ll keep working on this project and updating it with additional metrics, improvements, and better monitoring support as I learn more about the DGX Spark.
