NVIDIA GPU — DCGM Ultimate Dashboard

A Grafana dashboard for NVIDIA GPUs monitored purely through DCGM Exporter: no nvidia-smi-based exporter, no Kubernetes dependency. 89 panels built from a 168-field custom DCGM counter set, with full parity with 14574 wherever DCGM has an equivalent, plus everything nvidia-smi cannot give you. Scales from one workstation GPU to a multi-host NVLink/MIG/vGPU fleet.

Full documentation, more screenshots and source: GitHub

Key Features

• Fleet at a glance: GPUs monitored, exporters down, GPUs busy, avg util/temp, total power, Xid count, health, throttled-now. A worst-first Fleet Live Status table shows which GPU needs attention.
• Real activity (PROF_*): SM, tensor, DRAM and graphics-engine activity, FP64/FP32/FP16/INT precision mix, and estimated TFLOPS and bandwidth. These show how busy the GPU really is, beyond the headline GPU util %.
• Clocks and throttling: full decode of the clock-event reasons (why clocks are held down), with real throttling separated from informational limits. Also P-State history and time throttled per reason.
• Power, energy and cost: the full power-limit stack, energy used over the selected range, estimated cost and cost per hour, and idle-power waste.
• PCIe and NVLink: true byte counters instead of an estimate, link gen/width pass/fail, and replay/error rates.
• Reliability (ECC, Xid, health): ECC errors, row remapping, Xid errors by code, and the GPU recovery action. A per-subsystem DCGM health timeline includes a drill-down showing why a check went Warn or Fail.
• Datacenter row (collapsed by default): Fabric Manager/IMEX/C2C, MIG, vGPU license, Blackwell power profiles and NVLink health.

Requirements

• Grafana 10.4+, Prometheus, dcgm-exporter (verified on Grafana 13.2.2, Prometheus 3.14, dcgm-exporter 4.8.4 / DCGM 4.7).
• custom-counters.csv is required. It replaces the whole default field list, so it is not an add-on to the stock CSV.

Quick Setup

1. Download the CSV to the exporter host:

   curl -o /etc/dcgm-exporter/custom-counters.csv https://raw.githubusercontent.com/Axiumine/Grafana-NVIDIA-GPU-DCGM-Ultimate-Dashboard/main/custom-counters.csv
   

2. Run dcgm-exporter with it (systemd: systemctl edit nvidia-dcgm-exporter.service and override ExecStart; Docker/Kubernetes: mount the CSV and pass the same arguments):

   dcgm-exporter -f /etc/dcgm-exporter/custom-counters.csv -c 5000
   

3. Scrape it from Prometheus (scrape interval at least the exporter's -c interval):

   scrape_configs:
     - job_name: dcgm
       scrape_interval: 5s
       static_configs:
         - targets:
             - exporter-host:9400
   

4. Import the dashboard and pick your Prometheus datasource at the DS_PROMETHEUS prompt.

Variables

For the row-by-row panel list, why DCGM beats nvidia_gpu_exporter, DCGM gotchas and known limitations, see the README.

License: GPL-3.0-or-later (details).