Skip to content

Optional GPU telemetry metrics

GET /metrics under the existing read scope may include the following GPU observations. This additive contract serves hardware dashboards. It does not change capacity, pressure, admission, owner policy or authorization.

The NVIDIA capacity probe owns sampling. Sensor fields extend its existing read-only nvidia-smi query; HTTP scrapes only encode cached observations. A driver that rejects additional query fields falls back to the original capacity query within the existing five-second budget, then keeps that legacy query for the probe's lifetime. No additional polling process, GPU allocation, driver installation or temperature control is introduced.

Metric Labels Meaning and valid range
vacancyd_gpu_info resource, uuid, pci_bus_id, name Value 1 for the selected physical adapter. PCI is canonical 0000:01:00.0; UUID and product name are strings. Driver index is not an identity.
vacancyd_gpu_core_temperature_celsius resource Core temperature, 0–150 degrees Celsius.
vacancyd_gpu_memory_temperature_celsius resource Qualified memory temperature, greater than 0 through 150 degrees Celsius. Zero is unsupported, not a reading.
vacancyd_gpu_power_watts resource Current power, 0–2,000 watts.
vacancyd_gpu_fan_speed_percent resource Reported fan percentage, 0–100. Zero is valid.
vacancyd_gpu_sm_clock_megahertz resource SM clock, 0–10,000 MHz.
vacancyd_gpu_memory_clock_megahertz resource Memory clock, 0–40,000 MHz.
vacancyd_gpu_memory_utilization_percent resource Memory activity, 0–100 percent; independent from allocated bytes.

Unavailable, missing, negative, nonfinite, out-of-range or N/A values produce no series. Never encode unsupported sensors as zero. A newly missing value or failed capacity observation removes its previous sensor series. Consumers must also enforce the existing capacity observation age and Prometheus scrape age; scrape time is not a hardware timestamp. No sensor availability can authorize lending or invalidate otherwise usable capacity.

Identity is selected by the same configured selector as capacity, and old identity labels are removed on replacement or failure. Only one identity is exported per resource. Non-GPU probes produce no GPU series. There are no model, process, account, request, prompt or device-control labels.

Existing consumers may ignore these metrics. New consumers must tolerate their absence on older daemons and unsupported devices. No hotspot metric is promised: the driver query does not supply it. Host credentials, endpoint and port remain unchanged. Provider tests verify actual Prometheus encoding and missing/error transitions; dashboard fixtures use these exact names and units.