Nvidia Status

NVIDIA GPU dashboard + exporter — shows the GDDR6 / GDDR6X VRAM & junction temps that nvidia-smi hides on consumer RTX 30/40 cards (read straight off the GPU's PCIe BAR0), plus detailed stats including per‑Docker‑container GPU usage.

Exports to Prometheus, a JSON API, MQTT (with Home Assistant auto‑discovery) and InfluxDB (v1/v2) out of the box — one small container for live viewing and long‑term analysis in Grafana.

Built for Unraid, works on any Linux host with the NVIDIA proprietary driver.

Why?

nvidia-smi cannot report memory‑junction temperature on consumer GeForce cards — yet that's the number that actually throttles a 3090/4090 (GDDR6X junctions run 20–30 °C hotter than the core and throttle around 95–110 °C). Nvidia Status reads it directly from the memory‑controller registers, alongside everything nvidia-smi does expose, and presents it all in one place — plus which Docker container is using the GPU and what codec it's encoding.

Features

• Hidden temps: core + junction + VRAM (via BAR0) + all thresholds
• Power: draw, limit, min/max/default, and optional per‑GPU power‑limit control with a latch that re‑applies after a restart
• Clocks / utilisation / VRAM / fan / PCIe gen & width + live upload/download bandwidth
• Throttle reasons (power cap, thermal, power brake, …) as clear indicators
• Per‑process → container: each GPU process' VRAM / compute / encode / decode, mapped to the owning Docker container
• NVENC sessions: codec (H.264 / HEVC / AV1), resolution, fps — per session, per container
• Multi‑GPU aware throughout (per‑GPU cards, metrics, MQTT devices, InfluxDB rows, control)
• Five outputs: dashboard, Prometheus, JSON, MQTT/HA, InfluxDB — enable/disable each
• Built‑in Settings page — configure MQTT/InfluxDB and toggle features from the browser

GPU support

Everything except junction/VRAM temps works on any GPU the NVIDIA Linux driver supports. The BAR0 junction/VRAM read is architecture‑specific:

On unsupported cards junction/VRAM simply show blank — everything else still works.

Install (Unraid)

1. Install the Nvidia‑Driver plugin (Community Applications) if you haven't.
2. Search Community Applications for Nvidia Status and install, or add this repo's template.
3. Open the WebUI. Configure MQTT/InfluxDB from the Settings page if you want them.

docker run

docker run -d --name nvidia-status \
  --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all -e NVIDIA_DRIVER_CAPABILITIES=all \
  --cap-add SYS_RAWIO --device /dev/mem \
  --pid=host --cgroupns=host \
  -v /var/lib/docker/containers:/hostcontainers:ro \
  -v /mnt/user/appdata/nvidia-status:/config \
  -p 9835:9835 --restart unless-stopped \
  edddeduck/nvidia-status:latest

Add -e ENABLE_POWER_LIMIT=1 --privileged to enable power‑limit control.

Endpoints

Integrations

Configure from the Settings page or via env vars.

• Prometheus / Grafana / Netdata / Telegraf → scrape /metrics
• Home Assistant → set MQTT_HOST (+ user/pass); sensors auto‑appear via MQTT discovery, grouped as one device per GPU — including the hidden junction temp
• InfluxDB → Grafana → set INFLUX_URL (+ v1 INFLUX_DB or v2 INFLUX_TOKEN); writes line protocol directly, no Telegraf needed
• Homepage / Dashy / Node‑RED → read /json

Configuration

Everything has an env var (see the Unraid template / nvidia-status.xml), and can also be set on the Settings page (persisted to /config/settings.json, which overrides env vars). Key ones: NODE_NAME, INTERVAL, ENABLE_DASHBOARD/METRICS/JSON/MQTT/INFLUX, MQTT_HOST/PORT/USER/PASS, INFLUX_URL/DB/USER/PASS (or v2 INFLUX_TOKEN/ORG/BUCKET), ENABLE_POWER_LIMIT, PL_MIN/PL_MAX.

Security

• Reads the GPU's BAR0 via SYS_RAWIO + /dev/mem — required to read the junction/VRAM temps that Nvidia doesn't expose. This grants the container raw physical‑memory read.
• Per‑process → container names need --pid=host + --cgroupns=host + the read‑only /var/lib/docker/containers mount.
• Power‑limit control additionally requires --privileged and is off by default.
• Secrets you enter (MQTT/InfluxDB passwords) are stored in the container's /config volume.

Credits

• VRAM‑junction/hotspot reader: gputemps by ThomasBaruzier (Apache‑2.0) — vendored as gputemps.c, see GPUTEMPS-LICENSE.

License

MIT — see LICENSE. The vendored gputemps.c retains its Apache‑2.0 license.