Monitoring a DGX Station GB300 builds on what is already installed. NVIDIA pre-installs Data Center GPU Manager (DCGM), so the job is to add NVIDIA's DCGM-Exporter, which publishes GPU metrics on port 9400. Prometheus scrapes them, and a Grafana dashboard displays them. Add the BMC's out-of-band telemetry and your UPS's network card for the parts of the box the GPU can't see. Two warnings shape the whole setup. First, pin the exporter version, because the metric names on DCGM-Exporter's development branch are being renamed and will break dashboards and alerts when that release ships. Second, don't copy the exporter's documentation quickstart, which still uses a 2.1.4 image tagged for Ubuntu 20.04. This guide is compiled from NVIDIA's documentation and GitHub repositories as of 9 Nov 2026. SourceIT has not deployed it on a unit itself.
Which metrics matter most on a GB300?
| What to watch | Metric in DCGM-Exporter 4.8.4 defaults | Why it matters on a Station |
|---|---|---|
| Hardware errors | DCGM_FI_DEV_XID_ERRORS | XID errors are the GPU's own fault reports and should page someone |
| Memory health | DCGM_FI_DEV_UNCORRECTABLE_REMAPPED_ROWS, DCGM_FI_DEV_ROW_REMAP_FAILURE | Early warning on the 252 GB HBM3e before a hard failure |
| Temperature | DCGM_FI_DEV_GPU_TEMP, DCGM_FI_DEV_MEMORY_TEMP | The liquid loop is sealed; rising temperatures can point to a cooling problem |
| Power and clocks | DCGM_FI_DEV_POWER_USAGE, DCGM_FI_DEV_SM_CLOCK | The GB300 shares a power budget with any RTX PRO card, so caps show up as lower clocks |
| HBM in use | DCGM_FI_DEV_FB_USED, DCGM_FI_DEV_FB_FREE | Shows how close model servers are to the HBM limit |
| Real work | DCGM_FI_PROF_GR_ENGINE_ACTIVE, DCGM_FI_PROF_PIPE_TENSOR_ACTIVE | Better than plain utilisation for judging whether the GPU is earning its keep |
| Basic utilisation | DCGM_FI_DEV_GPU_UTIL | Simple busy or idle signal for capacity reports |
| Driver version | DCGM_FI_DRIVER_VERSION | Confirms updates landed as planned |
Source: etc/default-counters.csv at tag 4.8.4 in github.com/NVIDIA/dcgm-exporter (fetched 9 Nov 2026); NVIDIA DGX Station Development Guide, Dynamic Power Sloshing; DCGM-Exporter documentation (updated 27 May 2026). Metric names change on the exporter's main branch, as explained below.
What monitoring is already on the Station?
NVIDIA's DGX Station software stack page lists DCGM as pre-installed, alongside nvidia-smi for basic health checks. It also lists out-of-band telemetry "made available through the Baseboard Management Controller", through a web UI and Redfish. DCGM 4.6's documentation lists support for aarch64 and Ubuntu 24.04 LTS. The DCGM changelog adds GB300 devices "including workstation variants" in release 4.4.2, though it doesn't say outright that this means the DGX Station. DCGM can also report on the Grace CPU. Release 4.6.0 renamed those fields to forms such as DCGM_FI_DEV_CPU_TEMP_CELSIUS and DCGM_FI_DEV_CPU_POWER_WATTS. These CPU fields are not in the exporter's default counter list, so you would have to add them yourself.
Start by confirming DCGM is running and can see the GPU, using the commands from NVIDIA's DCGM Getting Started guide:
$ sudo systemctl --now enable nvidia-dcgm
$ dcgmi discovery -lSource: DCGM Getting Started. The guide also shows the datacenter-gpu-manager-4-cuda package for reinstalls; on a new Station you shouldn't need it.
How do you add DCGM-Exporter?
DCGM-Exporter reads from DCGM and exposes a Prometheus /metrics endpoint, listening on port 9400 by default. The latest release on GitHub at the time of writing is 4.8.4, built with DCGM 4.6.1. Pin that exact version in your container or service definition rather than tracking "latest". The release notes don't state which arm64 image tags exist, so confirm that the image you pull includes an arm64 build before relying on it. The Station's Grace CPU is Arm, and an x86-only image won't run. Once the exporter is running, NVIDIA's documentation checks it with:
$ curl localhost:9400/metricsSource: NVIDIA DCGM-Exporter documentation. Two recent changes can catch you out. Release 4.8.4 removed retired PCIe throughput fields from the supplied configurations, and NVIDIA warns that custom configurations still using them will fail. The previous release renamed the Hostname label to lowercase hostname, which affects any dashboard query that filters on it.
Pin the exporter version and write your alerts against the names that version emits. When the renamed metrics ship, update dashboards and alerts together, in one change window.
Why will your dashboards need updating?
The default-counters.csv file on DCGM-Exporter's main branch, which was unreleased on 9 Nov 2026, already uses unit-suffixed names. Examples include DCGM_FI_DEV_GPU_UTIL_RATIO, DCGM_FI_DEV_GPU_TEMP_CELSIUS, DCGM_FI_DEV_BOARD_POWER_WATTS, DCGM_FI_DEV_FB_USED_BYTES and DCGM_FI_PROF_TENSOR_UTIL_RATIO. It also adds commented-out exporter health fields such as DCGM_EXP_GPU_HEALTH_STATUS. Any Grafana panel or Prometheus alert written against the 4.8.4 names will stop showing data when you upgrade to a release that uses the new ones. Keep a list of every query that names a metric, and re-test them as part of each exporter upgrade.
How do Prometheus and Grafana fit in?
Prometheus needs one scrape job pointing at the Station's address on port 9400. Write it from Prometheus's own configuration documentation; NVIDIA's exporter page doesn't include one, and we won't guess at one here. For Grafana, the DCGM-Exporter README points to dashboard 12239 on grafana.com ("NVIDIA DCGM Exporter dashboard") and a grafana/dcgm-exporter-dashboard.json file in the repository. Check that the dashboard's queries use the metric names your pinned exporter emits. A dashboard built for one naming scheme shows empty panels under the other.
If the Station is part of a Kubernetes cluster, the exporter normally runs through NVIDIA's GPU Operator instead. Our Kubernetes integration guide covers that path.
What do power sharing and MIG change?
The GB300 and any RTX PRO card share the Station's power through NVIDIA's dynamic power sloshing. NVIDIA's Development Guide gives the GB300's cap as 1,600 W with no RTX card or one of 75 W or less, 1,250 W with a 76–300 W card, and 950 W with a card above 300 W. The tunables live in /etc/default/vsloshd. In practice, a lower power reading and lower SM clocks under load may mean the budget is working as designed, not that there is a fault. Note which display card is fitted before you set power alerts.
If you split the GPU with MIG, the exporter labels metrics with the instance profile and ID (GPU_I_PROFILE, GPU_I_ID), so you can chart each team's slice. NVIDIA's MIG guide adds two limits. "Profiling of shared GPU resources is not supported" under MIG. And processes holding the GPU, including DCGM, must be stopped before MIG is enabled. Our post on sharing one GB300 between seven teams covers the partitioning itself.
What about the parts of the box the GPU can't see?
Fans, power supply and board-level sensors sit behind the BMC, which publishes them through Redfish and its web interface. Exactly which sensors are exposed varies by OEM, so check your model's BMC documentation. Mains power is the other blind spot. A UPS with a network management card can raise power events and trigger a clean shutdown, and our PDU monitoring guide explains the same idea at rack level. Finally, decide who receives each alert and when. The who manages a DGX Station post covers ownership, and the post-delivery checklist makes joining monitoring part of handover. For the Station's first boot and GPU checks, start with our setup guide.
Buying in Singapore
For power telemetry, the APC UPS Network Management Card 3 is S$596.25 and the APC Smart-UPS SRT 3000VA online tower is S$4,165.00 (listed 9 Nov 2026). Size any UPS on watts against the Station's 1,600 W supply and check the UPS datasheet. Prices change; ask for a current quotation. The six DGX Station GB300 systems we list, from ASUS, Dell, HP, MSI, GIGABYTE and Supermicro, are price on request; start with the MSI XpertStation WS300 or the HP ZGX Fury G1n. Each is supplied through its OEM's channel. To discuss a configuration, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Is DCGM already installed on a DGX Station GB300? Yes. NVIDIA's software stack page lists Data Center GPU Manager as pre-installed. You add DCGM-Exporter, Prometheus and Grafana on top.
Which port does DCGM-Exporter use? Port 9400 by default. NVIDIA's documentation checks it with curl localhost:9400/metrics.
Can I copy the quickstart from NVIDIA's exporter docs? No. It still uses a 2.1.4 image tagged for Ubuntu 20.04. Pin a current release, such as 4.8.4, and confirm it has an arm64 build.
Why did my Grafana panels go blank after an upgrade? Check for renamed metrics or labels. A recent release renamed the Hostname label, and the exporter's main branch moves to unit-suffixed names such as DCGM_FI_DEV_GPU_UTIL_RATIO.
Can I see per-team usage when the GPU is split with MIG? Yes. The exporter labels metrics with the MIG instance profile and ID, though NVIDIA says profiling of shared resources isn't supported under MIG.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
