PANTAW COLLECTOR · GPU

nvidia collector

Distinguish NVIDIA hardware presence, active kernel driver and proprietary telemetry availability.

Source and collected facts

Read-only PCI/sysfs inventory plus an optional fixed nvidia-smi telemetry query through the shared command helper.

PCI address, vendor and device
bound kernel driver
nouveau and proprietary-driver state
nvidia-smi and telemetry availability
temperature, power, memory and utilization when available

Observation contract

Working proprietary telemetry produces SUCCESS. Hardware under nouveau, another driver or unknown driver state produces PARTIAL. No detected hardware or unusable evidence produces UNAVAILABLE according to the observed facts.

Relevant ReasonCode values: NVIDIA_OBSERVED, NVIDIA_PARTIAL, INTERNAL_ERROR

All collectors use structured ObservationStatus values and retain unavailable optional fields explicitly.

Graceful unavailability and limits

Proprietary metrics require a working NVIDIA driver and nvidia-smi. Hardware under nouveau remains detected rather than being described as absent.

Permissions and privacy

Permissions. PCI/sysfs reads are usually unprivileged; nvidia-smi access follows driver and device permissions.

Privacy. GPU UUID and PCI topology can identify hardware; public examples omit real UUIDs.

Sanitized example

{
  "source": "nvidia",
  "status": "SUCCESS",
  "severity": "INFO",
  "reason_code": "NVIDIA_OBSERVED",
  "target": "gpu:0000:01:00.0",
  "metadata": {"hardware_detected":true,"kernel_driver":"nouveau","nouveau_active":true,"nvidia_smi_available":false,"telemetry_available":false}
}

Synthetic example. It is not production node data.