nvidia-smi-exporter

A Prometheus exporter for NVIDIA GPU metrics via nvidia-smi. Compatible with all GPU architectures (Turing, Ampere, Ada Lovelace, Blackwell) and all driver versions, including driver 595+ which deprecated several fields across all architectures.

Why this project exists

The original exporter (and its forks) fail on driver 595+. Starting with driver 595, NVIDIA changed several nvidia-smi -q -x fields from N/A to the string "Requested functionality has been deprecated" — on all GPU architectures, not just Blackwell. The original filterNumber() function strips non-numeric characters from this string, producing an empty string, which is then emitted as a Prometheus metric value — invalid text exposition format that causes the entire scrape to fail silently, dropping all metrics for that interval.

This rewrite fixes that at the structural level: every value passes through a safe filter that always returns a valid numeric string or "0", and writeMetric refuses to emit any line with an empty value.

Credits

This project is derived from three MIT-licensed upstream projects:

• kristophjunge/docker-prometheus-nvidiasmi — original exporter
• e7d/docker-prometheus-nvidiasmi — fork adding gpu_power_readings support
• ich777/unraid-prometheus_nvidia_smi_exporter — Unraid packaging

See LICENSE for full attribution.

Docker Usage

Requirements: NVIDIA Container Toolkit installed on the host.

docker-compose

services:
  nvidia-smi-exporter:
    image: ccmpbll/nvidia-smi-exporter:latest
    runtime: nvidia
    ports:
      - "9202:9202/tcp"
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
    restart: unless-stopped

The container image includes nvidia-smi from the CUDA base image — no host bind-mount is needed. The NVIDIA runtime provides GPU device access at runtime.

docker run

docker run -d \
  --runtime=nvidia \
  -p 9202:9202 \
  -e NVIDIA_VISIBLE_DEVICES=all \
  ccmpbll/nvidia-smi-exporter:latest

Environment variables

Unraid Installation

Prerequisites

Install the Nvidia-Driver plugin from Community Applications first. This plugin installs nvidia-smi at /usr/bin/nvidia-smi on the Unraid host.

The nvidia-smi-exporter plugin will abort installation with a clear error message if /usr/bin/nvidia-smi is not found.

Install via Community Applications

Search for "nvidia-smi-exporter" in the Community Applications store, or add the plugin URL directly in the Unraid Plugins tab:

https://github.com/ccmpbll/nvidia-smi-exporter/raw/main/unraid/nvidia-smi-exporter.plg

Configuration

After installation, navigate to Settings → Nvidia SMI Exporter in the Unraid web UI to:

• Start / stop / restart the exporter
• Configure the listen port (default: 9202)
• View current GPU information

Settings are persisted to /boot/config/plugins/nvidia-smi-exporter/nvidia-smi-exporter.cfg.

Grafana Dashboard

A pre-built dashboard is included at grafana/dashboard.json.

Sections:

• Overview — attached GPU count, driver/CUDA version, performance state, fan speed, GPU temperature
• Utilization — GPU, memory controller, encoder, decoder, JPEG, and OFA utilization over time
• Memory — framebuffer and BAR1 memory (total / used / free) over time
• Temperature — GPU and memory temperature with threshold reference lines (T-Limit for Blackwell, max/slow thresholds for older GPUs)
• Power — power draw (with average and instantaneous for Blackwell), power limits (current, default, enforced, min, max)
• Clocks — current graphics, SM, memory, and video clock frequencies; max frequency stats
• PCIe — TX/RX throughput, PCIe generation and link width stats, replay counters
• Encoder & FBC — encoder and frame buffer capture session counts, average FPS, average latency
• Processes — per-process GPU memory usage over time

To import:

1. In Grafana, go to Dashboards → Import
2. Upload grafana/dashboard.json or paste its contents
3. Select your Prometheus datasource when prompted

The dashboard includes a GPU variable to filter by GPU name, supporting multi-GPU systems.

Prometheus Scrape Configuration

scrape_configs:
  - job_name: nvidia_smi
    static_configs:
      - targets:
          - 'your-host:9202'
    scrape_interval: 15s

Metrics Reference

All metrics carry these labels:

nvidiasmi_process_used_memory_bytes additionally carries: process_pid, process_name, process_type.

System

PCIe

GPU State

Memory

Utilization

Encoder / FBC Stats

Temperature

Power

Clocks

Processes

Additional labels: process_pid, process_name, process_type (C=compute, G=graphics).

Known Limitations

Driver 595+ (all architectures)

Starting with driver 595, several nvidia-smi fields return "Requested functionality has been deprecated" rather than a numeric value or N/A, on all GPU architectures. This exporter handles them as follows:

P-state metrics

nvidiasmi_power_state_int and nvidiasmi_gpu_power_state_int are emitted only when the source value is a valid P-state string (e.g. P0, P8). They are silently omitted on driver 595+ where the field is deprecated. Pre-existing dashboards that graph these metrics will show gaps on affected GPUs rather than incorrect zero values.

Building from Source

# Docker image
docker build -f docker/Dockerfile -t nvidia-smi-exporter .

# Native binary (requires Go 1.22+)
cd src && go build -o nvidia-smi-exporter .

# Test with sample XML
TEST_MODE=1 ./nvidia-smi-exporter
curl -s http://localhost:9202/metrics

# Unraid package
./unraid/build.sh 2026.05.12

Contributing

Issues and PRs welcome. When reporting a metrics bug, please include the output of nvidia-smi -q -x from your system (you can redact UUIDs and serial numbers).