windows_exporter/docs/collector.gpu.md

GPU collector

The gpu collector exposes metrics about GPU usage and memory consumption, both at the adapter (physical GPU) and per-process level.

Metric name prefix	gpu
Data source	Perflib, D3DKMT (gdi32.dll), Configuration Manager (cfgmgr32.dll)
Counters	GPU Engine, GPU Adapter Memory, GPU Local Adapter Memory, GPU Non Local Adapter Memory, GPU Process Memory
Enabled by default?	No

GPU adapters are discovered through the D3DKMT API. Software adapters like the Microsoft Basic Render Driver are not exposed. If an adapter can't be fully discovered, a warning is logged and the collector keeps working with the remaining adapters. The list of adapters is refreshed when a performance counter references an unknown adapter, e.g. after a driver update, a device restart or when an external GPU is attached (at most once every 30 seconds).

Flags

None

Metrics

These metrics are available on supported versions of Windows with compatible GPUs and drivers:

Adapter-level Metrics

Name	Description	Type	Labels
windows_gpu_info	A metric with a constant '1' value labeled with GPU device information.	gauge	architecture,bus_number,device_id,device_number,driver_version,function_number,luid,name,wddm_version
windows_gpu_dedicated_system_memory_size_bytes	The size, in bytes, of memory that is dedicated from system memory.	gauge	device_id,luid
windows_gpu_dedicated_video_memory_size_bytes	The size, in bytes, of memory that is dedicated from video memory.	gauge	device_id,luid
windows_gpu_shared_system_memory_size_bytes	The size, in bytes, of memory from system memory that can be shared by many users.	gauge	device_id,luid
windows_gpu_adapter_memory_committed_bytes	Total committed GPU memory in bytes per physical GPU	gauge	device_id,luid,phys
windows_gpu_adapter_memory_dedicated_bytes	Dedicated GPU memory usage in bytes per physical GPU	gauge	device_id,luid,phys
windows_gpu_adapter_memory_shared_bytes	Shared GPU memory usage in bytes per physical GPU	gauge	device_id,luid,phys
windows_gpu_local_adapter_memory_bytes	Local adapter memory usage in bytes per physical GPU	gauge	device_id,luid,phys,part
windows_gpu_non_local_adapter_memory_bytes	Non-local adapter memory usage in bytes per physical GPU	gauge	device_id,luid,phys,part

The driver_version, wddm_version and architecture labels of windows_gpu_info are empty if the driver does not report them.

Sensor Metrics

The sensor metrics are read from the graphics kernel (dxgkrnl) with D3DKMTQueryAdapterInfo, the same source Task Manager uses. They are vendor-neutral, but depend on WDDM 2.4 or newer and on driver support. Drivers report 0 for values they do not support, so a metric is only exposed if the driver reported a non-zero value or capability for it when the adapter was discovered. Software adapters, like the Microsoft Basic Render Driver, are skipped and never queried. Query failures are logged at debug level and never fail the scrape.

Name	Description	Type	Labels
windows_gpu_temperature_celsius	Main temperature sensor reading of the physical GPU in degrees Celsius	gauge	device_id,luid,phys
windows_gpu_temperature_warning_celsius	Temperature in degrees Celsius at which the physical GPU starts throttling	gauge	device_id,luid,phys
windows_gpu_temperature_max_celsius	Maximum temperature in degrees Celsius before the physical GPU takes damage	gauge	device_id,luid,phys
windows_gpu_fan_speed_rpm	Current speed of the main fan in revolutions per minute (0 while the fan is stopped)	gauge	device_id,luid,phys
windows_gpu_fan_speed_max_rpm	Maximum speed of the main fan in revolutions per minute	gauge	device_id,luid,phys
windows_gpu_power_usage_ratio	Current power draw as a ratio (0–1) of the maximum power (TDP) of the physical GPU	gauge	device_id,luid,phys
windows_gpu_memory_frequency_hertz	Current clock frequency of the GPU memory in hertz	gauge	device_id,luid,phys
windows_gpu_memory_frequency_max_hertz	Maximum clock frequency of the GPU memory in hertz, while not overclocked	gauge	device_id,luid,phys
windows_gpu_engine_frequency_hertz	Current clock frequency of the GPU engine in hertz	gauge	device_id,luid,phys,eng
windows_gpu_engine_frequency_max_hertz	Maximum clock frequency of the GPU engine in hertz, while not overclocked	gauge	device_id,luid,phys,eng

The engine frequency metrics are only exposed for engines that report a maximum frequency, typically the 3D engine. The eng label matches the eng label of windows_gpu_engine_time_seconds.

Per-process Metrics

Name	Description	Type	Labels
windows_gpu_engine_time_seconds	Total running time of the GPU engine in seconds	counter	device_id,luid,phys, eng, engtype, process_id
windows_gpu_process_memory_committed_bytes	Total committed GPU memory in bytes per process	gauge	device_id,luid,phys,process_id
windows_gpu_process_memory_dedicated_bytes	Dedicated GPU memory usage in bytes per process	gauge	device_id,luid,phys,process_id
windows_gpu_process_memory_local_bytes	Local GPU memory usage in bytes per process	gauge	device_id,luid,phys,process_id
windows_gpu_process_memory_non_local_bytes	Non-local GPU memory usage in bytes per process	gauge	device_id,luid,phys,process_id
windows_gpu_process_memory_shared_bytes	Shared GPU memory usage in bytes per process	gauge	device_id,luid,phys,process_id

Metric Labels

• luid: Locally unique identifier of the GPU adapter (e.g., "0x00000000_0x0001136A"), formatted as <HighPart>_<LowPart>. It is assigned by Windows and identifies the adapter in all metrics of this collector. Use it to join metrics with windows_gpu_info. The LUID is not stable: it changes when the driver is updated, the device is restarted or the system reboots.
• device_id: PnP device instance ID of the GPU adapter (e.g., "PCI\VEN_10DE&DEV_1B81&SUBSYS_61733842&REV_A1\4&1d81e16&0&0019"). If the device instance can't be resolved, the PnP device ID without the instance part is used.
• phys: Physical adapter index within a (linked) logical adapter (e.g., "0")
• bus_number,device_number,function_number: PCI address of the GPU adapter (windows_gpu_info only)
• part: Memory segment index (e.g., "0")
• eng: GPU engine index (e.g., "0", "1", ...)
• engtype: GPU engine type (e.g., "3D", "Copy", "VideoDecode", etc.)
• process_id: Process ID
• driver_version: Kernel mode driver version (e.g., "32.0.15.8266")
• wddm_version: WDDM version of the driver (e.g., "3.2")
• architecture: GPU architecture as reported by the driver (e.g., "Pascal")

Example Metric

These are basic queries to help you get started with GPU monitoring on Windows using Prometheus.

Show GPU information:

windows_gpu_info{architecture="Pascal",bus_number="8",device_id="PCI\\VEN_10DE&DEV_1B81&SUBSYS_61733842&REV_A1\\4&1d81e16&0&0019",device_number="0",driver_version="32.0.15.8266",function_number="0",luid="0x00000000_0x0001136A",name="NVIDIA GeForce GTX 1070",wddm_version="3.2"} 1

Add the GPU name to a metric:

windows_gpu_adapter_memory_dedicated_bytes * on(luid) group_left(name) windows_gpu_info

Show total dedicated GPU memory (in bytes) usage on GPU 0:

windows_gpu_adapter_memory_dedicated_bytes{phys="0"}

Aggregate GPU utilization across all processes for a physical GPU (3D engine):

sum by (phys) (
  rate(windows_gpu_engine_time_seconds{phys="0", engtype="3D"}[1m])
) * 100

Show GPU utilization for a specific process (3D engine):

sum by (phys, process_id) (
  rate(windows_gpu_engine_time_seconds{process_id="1234", engtype="3D"}[1m])
) * 100

Show dedicated GPU memory per process:

sum by (luid, process_id) (
  windows_gpu_process_memory_dedicated_bytes
)

Useful Queries

Show top 5 processes by GPU utilization (all engines):

topk(5, sum by (process_id) (
  rate(windows_gpu_engine_time_seconds[1m])
) * 100)

Show GPU memory usage per physical GPU:

sum by (phys) (
  windows_gpu_adapter_memory_dedicated_bytes
)

Show GPU engine time with process owner and command line:

windows_gpu_engine_time_seconds * on(process_id) group_left(owner, cmdline) windows_process_info

Show the engine clock as a ratio of its maximum clock:

windows_gpu_engine_frequency_hertz / windows_gpu_engine_frequency_max_hertz

Alerting Examples

prometheus.rules

# Alert on processes using more than 80% of a GPU's capacity over 10 minutes
- alert: HighGpuUtilization
  expr: |
    sum by (process_id) (
      rate(windows_gpu_engine_time_seconds[1m])
    ) * 100 > 80
  for: 10m
  labels:
    severity: warning
  annotations:
    summary: "High GPU Utilization (process {{ $labels.process_id }})"
    description: "Process is using more than 80% of GPU resources\n  VALUE = {{ $value }}\n  LABELS: {{ $labels }}"
# Alert on GPUs running within 5 °C of their throttling temperature
- alert: GpuTemperatureHigh
  expr: |
    windows_gpu_temperature_celsius > windows_gpu_temperature_warning_celsius - 5
  for: 5m
  labels:
    severity: warning
  annotations:
    summary: "GPU temperature high (luid {{ $labels.luid }})"
    description: "GPU is close to its throttling temperature\n  VALUE = {{ $value }}\n  LABELS: {{ $labels }}"

Notes

• Per-process metrics allow you to identify which processes are consuming GPU resources.
• Adapter-level metrics provide an overview of total GPU memory usage.
• For overall GPU utilization, aggregate per-process metrics in Prometheus using queries such as sum().
• The collector relies on Windows performance counters; ensure your system and drivers support these counters.
• Sensor metrics (temperature, fan, power, clocks) require WDDM 2.4 or newer and a driver that reports these values.

Enabling the Collector

To enable the GPU collector, add gpu to the list of enabled collectors in your windows_exporter configuration.

Example (command line):

windows_exporter.exe --collectors.enabled=gpu