Metrics Exporter Metrics Reference¶

This page lists every metric the RBLN Metrics Exporter exposes, with its labels and the version that introduced it.

The following metrics are exported for each NPU device, tagged with the device UUID, card name, and character device node (rblnN).

Name	Deprecated alias	Description	Unit	Minimum version
`rbln_npu_temperature`	`RBLN_DEVICE_STATUS:TEMPERATURE`	Temperature	°C	`v0.2.4`
`rbln_npu_power`	`RBLN_DEVICE_STATUS:CARD_POWER`	Power usage	W	`v0.2.4`
`rbln_npu_memory_used`	`RBLN_DEVICE_STATUS:DRAM_USED`	DRAM in use	Bytes	`v0.2.4`
`rbln_npu_memory_total`	`RBLN_DEVICE_STATUS:DRAM_TOTAL`	Total DRAM	Bytes	`v0.2.4`
`rbln_npu_utilization`	`RBLN_DEVICE_STATUS:UTILIZATION`	Utilization	%	`v0.2.4`
`rbln_npu_health`	`RBLN_DEVICE_STATUS:HEALTH`	NPU health status	0/1	`v0.2.4`
`rbln_npu_device_status`	N/A	Device state machine status (one series per `state` label; only the current state is 1)	0/1	`v0.3.0`
`rbln_npu_power_state`	N/A	DVFS performance state	Level	`v0.3.0`
`rbln_npu_pcie_link_speed_gts`	N/A	Current PCIe link speed	GT/s	`v0.3.0`
`rbln_npu_pcie_link_width`	N/A	Current PCIe link width	Lanes	`v0.3.0`
`rbln_npu_device_info`	N/A	Device identity and static attributes exposed as labels	Always 1	`v0.3.0`
`rbln_up`	N/A	Whether the last metrics collection from RSMD succeeded	0/1	`v0.3.1`
`rbln_npu_device_shared`	N/A	Whether the device's `ResourceClaim` is shared by more than one Pod, so its Kubernetes labels name only one of them	0/1	`v0.3.3`

Note

rbln_npu_health (deprecated alias: RBLN_DEVICE_STATUS:HEALTH) is a binary state metric. 0 means the NPU is active, while 1 means it is inactive.

Note

rbln_up is exposed as 0 between exporter start and the first collection cycle.

The exporter does not expose the following metrics when it cannot determine their values.

rbln_npu_power_state: devices whose daemon does not report a performance state value
rbln_npu_pcie_link_speed_gts, rbln_npu_pcie_link_width: devices whose topology information could not be queried

Common NPU Metrics Label Attributes¶

Label	Description
`name`	Character device node exposed by the kernel driver (`Device.name`, e.g., `rbln0`)
`uuid`	Globally unique identifier for the NPU device (`Device.uuid`)
`card`	Card product name reported in `Device.card_name` (e.g., `RBLN-CA25`)
`deviceID`	PCIe device ID reported in the proto (`Device.dev_id`, e.g., `1250`)
`hostname`	Name of the node the exporter runs on: `--node-name`, then the `NODE_NAME` environment variable, then the host name. In gateway mode, it is the host part of the target address.
`driver_version`	Kernel driver build returned by `VersionInfo.drv_version`
`firmware_version`	Firmware revision returned by `VersionInfo.fw_version`

Kubernetes NPU Metrics Label Attributes¶

Label	Description
`namespace`	Namespace for the workload using the device. Taken from `Pod.metadata.namespace`
`container`	Name of the container consuming the NPU. Taken from `Pod.spec.containers[].name`
`pod`	Name of the Pod holding the NPU allocation. Taken from `Pod.metadata.name`

Metric-specific Label Attributes¶

In addition to the common labels, rbln_npu_device_status uses a state label.

Label	Metric	Values
`state`	`rbln_npu_device_status`	`ready` · `busy` · `init` · `fault` · `finish` · `not_found`

Device Info Metric Label Attributes¶

In addition to the common labels, rbln_npu_device_info exposes the device's static attributes with the following labels.

Label	Description
`smc_version`	SMC firmware version
`pci_bus_id`	PCI bus address
`numa_node`	NUMA node number the device is attached to (`-1` when NUMA information is unavailable)
`rsd_group`	RSD group the device belongs to
`cpu_list`	Local CPU list of the device
`is_vf`	Whether the device is an SR-IOV virtual function (VF) (`true`/`false`)
`parent_name`	Parent physical device name when the device is a VF
`num_vfs`	Number of VFs configured on the physical device

Note

The exporter exposes the numa_node, rsd_group, and cpu_list labels as empty values when topology information cannot be queried.