Query

Print a report of the main attributes of each GPU, as an indented Key : Value listing: identity, PCI link, clock throttle reasons, memory, utilization, temperatures, power and clocks, and the processes using the device. Use Discovery for static detail such as compute resources and firmware components, and --query-gpu or Dump for any single metric. The report can be restricted to chosen sections, written as JSON, repeated at an interval, or saved to a file.

Synopsis

xpu-smi query
xpu-smi query --display [sections] --device [deviceIds]
xpu-smi query --json
xpu-smi query --loop [seconds] --count [count]
xpu-smi query --file [filename]

Options

-h, --help

Print this help message and exit.

-j, --json

Print the report as JSON instead of text. See JSON Output.

--verbose

Print every row, including those a device does not report, which are otherwise left out of the text and the JSON. A value the driver or OS says the device does not support reads Not Supported.

-d <deviceIds>, --device <deviceIds>, --id <deviceIds>

Device IDs or PCI BDF addresses to report on. Multiple IDs can be separated by a comma. Use -1 to select all devices. Without this option every device is reported.

--display <sections>

Comma-separated list of sections to print, from the table below. Names are case-insensitive. Sections always print in the order of the full report, whatever order they are listed in. Without this option everything is printed.

--loop <seconds>

Print the report again every <seconds> seconds until interrupted with Ctrl+C. On a terminal the whole report is drawn once and then only the lines whose values changed are rewritten, so unchanged data stays on screen. A line wider than the window wraps onto the rows below it. This needs the whole report, wrapped lines included, to fit in the window. If it does not, the report is printed once and a notice explains why, and nothing more is printed until the window is made large enough, at which point live updates start. Making the window smaller again prints the report once more. Use --display to select fewer sections if you want live updates in a smaller window. When output goes to a file or a pipe, each report is appended in full.

--loop-ms <milliseconds>

Like --loop, with the interval in milliseconds. Takes precedence over --loop.

--count <count>

Number of reports to print. Without --loop or --loop-ms a --count above 1 repeats once a second. A count of 1 prints a single report.

-f <path>, --file <path>

Write the report to <path> instead of the terminal. An existing file is overwritten.

Sections

--display accepts the names below. A field that belongs to a section is printed wherever it sits in the report, so --display PCI prints the PCI address fields from the PCI part of the report, and memory utilization appears under both MEMORY and UTILIZATION. A section or heading with nothing to show is left out. Without --display every section is printed except SUPPORTED_CLOCKS, which must be named (or ALL given) because it lists every clock the frequency range can be set to.

Section

Report content

IDENTITY

Product name and brand, stepping, index, DRM minor number, serial number, UUID, GFX firmware version, board part number, device state, whether the device is a physical or an SR-IOV virtual function, the recovery action it needs, and the PCI address

MEMORY

Memory total, used and free, memory type, number of channels and bus width, and memory utilization

UTILIZATION

GPU, memory, compute, render, media and copy engine utilization

TEMPERATURE

GPU core and memory temperatures and their configured shutdown and slowdown temperatures

POWER

Power draw and the current power limit

CLOCK

Current graphics, memory and media clocks, the graphics clock the driver requested (Applications Clocks), and the maximum graphics and media clocks

PCI

PCI address fields, class code and IDs, the physical slot, link generation and width (maximum and current), and throughput

EU_ARRAY

Nothing in this report. Use --query-gpu eu.active,eu.stall,eu.idle for the EU array figures

FAN

The fan speed, on devices with fans

COMPUTE

Scheduler mode, and the number of tiles, slices, sub-slices per slice and EUs

PERFORMANCE

Whether each clock throttle reason is active, the time throttled, and the performance factors, where the driver reports them

SUPPORTED_CLOCKS

The GPU core clocks the minimum and maximum of config --frequencyrange can be set to, summarised as low-high step N when evenly spaced. The memory clock cannot be set, so it has no list. Printed only when named

PIDS

Processes using the device

ALL

Everything above, including SUPPORTED_CLOCKS

Sample Output

From an Intel Arc B580 run by a non-root user.

==============XPU-SMI LOG==============

Timestamp                                 : 2026-10-23T14:50:00Z
UMD Version                               : 1.17.39980
KMD Version                               : N/A
Level Zero Version                        : 1.33.1

Attached GPUs                             : 1
GPU 0000:03:00.0
    Product Name                          : Intel(R) Arc(TM) B580 Graphics
    Product Brand                         : Intel(R) Corporation
    Stepping                              : A0
    GPU IP Version                        : 0x05004000
    Serial Number                         : unknown
    GPU UUID                              : 00000000-0000-0003-0000-0000e20b8086
    Device Index                          : 0
    Minor Number                          : 0
    GFX Firmware Version                  : N/A (requires read-write access to /dev/mei1)
    Platform Info
        Slot Number                       : N/A
    Device State                          : normal
    GPU Virtualization Mode
        Virtualization Mode               : physical
        Enabled VFs                       : 0
    GPU Recovery Action                   : None
    PCI
        Bus                               : 0x03
        Device                            : 0x00
        Domain                            : 0x0000
        Base Classcode                    : 0x03
        Sub Classcode                     : 0x00
        Device Id                         : 0xE20B
        Bus Id                            : 0000:03:00.0
        Sub System Id                     : 0x1100
        GPU Link Info
            PCIe Generation
                Max                       : N/A (requires CAP_SYS_ADMIN)
                Current                   : N/A (requires CAP_SYS_ADMIN)
                Host Max                  : 5
            Link Width
                Max                       : N/A (requires CAP_SYS_ADMIN)
                Current                   : N/A (requires CAP_SYS_ADMIN)
                Host Max                  : 16
        Tx Throughput                     : 102 kB/s
        Rx Throughput                     : 1623 kB/s
    Clocks Event Reasons
        Thermal Limit                     : Not Active
        Average Power Cap                 : Not Active
        Burst Power Cap                   : Not Active
        Current Limit                     : Not Active
        PSU Alert                         : Not Active
        SW Range                          : Not Active
        HW Range                          : Not Active
        Voltage                           : Not Active
        Thermal Alert                     : Not Active
        Power Alert                       : Not Active
    FB Memory Usage
        Total                             : 12216 MiB
        Reserved                          : 0 MiB
        Used                              : 218.62 MiB
        Free                              : 11997.38 MiB
    BAR Memory
        Total                             : 16384 MiB
    Memory Info
        Memory Type                       : GDDR6
        Number of Memory Channels         : 6
        Memory Bus Width                  : 192
    Scheduler Mode                        : timeslice
    Compute Resources
        Number of Tiles                   : 1
        Number of Slices                  : 5
        Number of Sub-slices per Slice    : 4
        Number of EUs                     : 160
    Utilization
        GPU                               : 0.00 %
        Memory                            : 1.79 %
        Compute                           : 0.00 %
        Render                            : 0.00 %
        Media                             : 0.00 %
        Video                             : 0.00 %
        Video Enhancement                 : 0.00 %
        Copy                              : 0.00 %
    Temperature
        GPU Current Temp                  : 30.00 C
        GPU Shutdown Temp                 : 130 C
        GPU Slowdown Temp                 : 105 C
        Memory Current Temp               : 48.00 C
        Memory Shutdown Temp              : 100 C
        Memory Slowdown Temp              : 85 C
    GPU Power Readings
        Average Power Draw                : 59.34 W
        Instantaneous Power Draw          : 68.62 W
        GPU Domain Power Draw             : 0.06 W
        Current Power Limit               : 380.00 W
        Peak Power Limit                  : 380.00 W
        Memory Power Draw                 : 8.38 W
    Clocks
        Graphics                          : 400 MHz
        Memory                            : 9500 MHz
        Media                             : 1200 MHz
    Applications Clocks
        Graphics                          : 400 MHz
    Max Clocks
        Graphics                          : 2850 MHz
        Media                             : 1500 MHz
    Processes
        None

Fields

Every row the report can print. Row is the label path in the text report, with the path to the value in --json output (under each entry of gpus) below it. Details gives the --display section that selects the row, its unit, and its source: the --query-gpu field that reads the same value, or the Discovery property it comes from. A row reads N/A when the device, driver or platform cannot supply its value; Meaning says when that is expected.

Report header

Printed once, above the devices. JSON keys are at the top level.

Row

Details

Meaning

Timestamp

timestamp

When sampling began, as ISO 8601 UTC.

UMD Version

umd_version

Source: driver_version

Level Zero user-mode GPU driver version.

KMD Version

kmd_version

Source: kernel_driver_version

Kernel-mode driver version that the driver module declares; N/A for a driver that declares none, such as the in-tree driver, and on Windows.

Level Zero Version

level_zero_version

Source: Level Zero loader

Version of the Level Zero loader.

Attached GPUs

attached_gpus

Number of devices in the report.

GPU <address>

pci_bus_id

Source: pci.bus_id

Starts each device; the value is its PCI address. In JSON each device is an entry of gpus.

Identity

Row

Details

Meaning

Product Name

product_name

Section: IDENTITY

Source: name

Device name the driver reports. Intel(R) Graphics [0x<device ID>] when the driver has no product name for the device.

Product Brand

product_brand

Section: IDENTITY

Source: discovery Vendor Name

Brand the driver reports, e.g. Intel(R) Corporation.

Stepping

stepping

Section: IDENTITY

Source: discovery Stepping

Silicon stepping, e.g. A0.

GPU IP Version

gpu_ip_version

Section: IDENTITY

Source: ip_version

Graphics IP version the driver reports, in hex (ze_device_ip_version_ext_t). It identifies the architecture: newer devices have higher values.

Serial Number

serial_number

Section: IDENTITY

Source: serial

Board serial number: from the AMC on a card that has one, otherwise the OEM serial number held by the GSC (root only). unknown when neither holds one.

GPU UUID

gpu_uuid

Section: IDENTITY

Source: uuid

Device UUID.

Device Index

device_index

Section: IDENTITY

Source: index

Index --device and the other commands accept for the device.

Minor Number

minor_number

Section: IDENTITY

Source: DRM device

Number of the device’s DRM card node, e.g. 1 for /dev/dri/card1. N/A on Windows.

GFX Firmware Version

gfx_firmware_version

Section: IDENTITY

Source: vbios_version

Version of the GFX firmware the device runs, read from the GSC. Needs read-write access to the device’s /dev/mei* node; without it the row says which node.

Board Part Number

board_part_number

Section: IDENTITY

Source: discovery Part Number

Part number of the board, read from the AMC. Present only on a device with an AMC.

Platform Info > Slot Number

platform_info.slot_number

Section: PCI

Source: discovery PCI Slot

Physical slot the platform firmware names for the device, e.g. GPU Board Slot2. N/A when the firmware names none, as on most desktop boards.

Device State

device_state

Section: IDENTITY

Source: device

normal, or survivability for a device whose firmware needs recovery.

GPU Virtualization Mode > Virtualization Mode

gpu_virtualization_mode.virtualization_mode

Section: IDENTITY

Source: PCI function

physical for a physical function, virtual for an SR-IOV virtual function.

GPU Virtualization Mode > Enabled VFs

gpu_virtualization_mode.enabled_vfs

Section: IDENTITY

Source: vf.enabled

SR-IOV virtual functions enabled on the physical function. Left out where the device or driver does not support SR-IOV.

GPU Recovery Action

gpu_recovery_action

Section: IDENTITY

Source: recovery_action

What the driver says the device needs to keep working: None, Reset (for example a wedged device) or Reset to repair. A device in survivability mode shows Firmware update and GPU reset.

Row

Details

Meaning

PCI > Bus

pci.bus

Section: IDENTITY, PCI

Source: pci.bus_id

Bus number of the PCI address, in hex.

PCI > Device

pci.device

Section: IDENTITY, PCI

Source: pci.bus_id

Device number of the PCI address, in hex.

PCI > Domain

pci.domain

Section: IDENTITY, PCI

Source: pci.bus_id

Domain of the PCI address, in hex.

PCI > Base Classcode

pci.base_classcode

Section: PCI

Source: pci.base_class

PCI base class, in hex: 0x03 display controller, 0x12 processing accelerator.

PCI > Sub Classcode

pci.sub_classcode

Section: PCI

Source: pci.sub_class

PCI subclass, in hex.

PCI > Device Id

pci.device_id

Section: IDENTITY, PCI

Source: pci.device_id

PCI device ID, in hex.

PCI > Bus Id

pci.bus_id

Section: IDENTITY, PCI

Source: pci.bus_id

Full PCI address, DDDD:BB:DD.F.

PCI > Sub System Id

pci.sub_system_id

Section: IDENTITY, PCI

Source: pci.sub_device_id

PCI subsystem device ID, in hex. N/A on Windows.

PCI > GPU Link Info > PCIe Generation > Max

pci.gpu_link_info.pcie_generation.max

Section: PCI

Source: pcie.link.gen.max

Highest PCIe generation the link supports. The driver reads it from PCI configuration space, which needs CAP_SYS_ADMIN; N/A when the driver does not know it.

PCI > GPU Link Info > PCIe Generation > Current

pci.gpu_link_info.pcie_generation.current

Section: PCI

Source: pcie.link.gen.current

PCIe generation the link runs at now. Needs CAP_SYS_ADMIN, as Max does.

PCI > GPU Link Info > PCIe Generation > Host Max

pci.gpu_link_info.pcie_generation.host_max

Section: PCI

Source: pcie.link.gen.host_max

Maximum PCIe generation of the root port the device is under, the CPU end of the path (Linux sysfs). A switch in between can limit the link further.

PCI > GPU Link Info > Link Width > Max

pci.gpu_link_info.link_width.max

Section: PCI

Source: pcie.link.width.max

Widest the link can be, in lanes. Needs CAP_SYS_ADMIN, as the generation does.

PCI > GPU Link Info > Link Width > Current

pci.gpu_link_info.link_width.current

Section: PCI

Source: pcie.link.width.current

Lanes the link uses now.

PCI > GPU Link Info > Link Width > Host Max

pci.gpu_link_info.link_width.host_max

Section: PCI

Source: pcie.link.width.host_max

Maximum PCIe link width of the root port the device is under (Linux sysfs).

PCI > PCIe Errors > Correctable

pci.pcie_errors.correctable

Section: PCI

Source: pcie.errors.correctable

PCIe correctable errors the device’s function reported since it was enumerated (AER, Linux). Left out where the function has no AER counters, as the B580 GPU function, whose switch ports carry them.

PCI > PCIe Errors > Non-Fatal

pci.pcie_errors.non_fatal

Section: PCI

Source: pcie.errors.non_fatal

PCIe uncorrectable non-fatal errors since enumeration (AER, Linux).

PCI > PCIe Errors > Fatal

pci.pcie_errors.fatal

Section: PCI

Source: pcie.errors.fatal

PCIe uncorrectable fatal errors since enumeration (AER, Linux).

PCI > PCIe Errors > Replay Number Rollovers

pci.pcie_errors.replay_number_rollovers

Section: PCI

Source: pcie.errors.replay_rollovers

Times the link’s REPLAY_NUM counter rolled over, a correctable link error (AER, Linux).

PCI > PCIe Errors > Replay Timer Timeouts

pci.pcie_errors.replay_timer_timeouts

Section: PCI

Source: pcie.errors.replay_timeouts

Times the link’s replay timer expired, a correctable link error (AER, Linux).

PCI > Tx Throughput

pci.tx_throughput

Section: PCI

Unit: kB/s

Source: pcie.tx.throughput.kbs

Data the device sent over PCIe, averaged over the sampling window. 0 on an idle link; N/A when the driver cannot read the PCIe counters.

PCI > Rx Throughput

pci.rx_throughput

Section: PCI

Unit: kB/s

Source: pcie.rx.throughput.kbs

Data the device received over PCIe, as for Tx Throughput.

PCI > Atomic Caps Inbound

pci.atomic_caps_inbound

Section: PCI

Source: pcie.atomic.completer

PCIe AtomicOp requests the device completes, from Device Capabilities 2: e.g. FETCHADD_32 SWAP_32 CAS_32, or None. Reading it needs CAP_SYS_ADMIN (Linux).

PCI > Replays Since Reset

pci.replays_since_reset

Section: PCI

Source: pcie.replay.counter

PCIe replays since the last reset, from Level Zero. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not. CRI’s driver has no replay counter; the PCIe Errors rows are the Linux alternative.

PCI > Atomic Requests

pci.atomic_requests

Section: PCI

Source: pcie.atomic.requests

Whether the device may send PCIe AtomicOp requests (Device Control 2): Enabled or Disabled. Needs CAP_SYS_ADMIN (Linux).

Clocks Event Reasons

Each row is Active when that reason holds the GPU clock back, otherwise Not Active. clocks.throttle.reason gives the active ones as a single list.

Row

Details

Meaning

Fan Speed

fan_speed

Section: FAN

Unit: %

Source: fan.speed

Fan speed, averaged across the device’s fans. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not.

Performance Factor > Compute

performance_factor.compute

Section: PERFORMANCE

Source: performance.factor.compute

How the hardware trades frequency toward the compute engines, 0-100; 50 is balanced. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not. xe disables it.

Performance Factor > Media

performance_factor.media

Section: PERFORMANCE

Source: performance.factor.media

The same for the media engines. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not.

PCI

Clocks Event Reasons > Thermal Limit

clocks_event_reasons.thermal_limit

Section: PERFORMANCE

Source: clocks.event_reasons.thermal_limit

The thermal limit.

Clocks Event Reasons > Average Power Cap

clocks_event_reasons.average_power_cap

Section: PERFORMANCE

Source: clocks.event_reasons.average_power_cap

The average (sustained) power limit.

Clocks Event Reasons > Burst Power Cap

clocks_event_reasons.burst_power_cap

Section: PERFORMANCE

Source: clocks.event_reasons.burst_power_cap

The burst power limit.

Clocks Event Reasons > Current Limit

clocks_event_reasons.current_limit

Section: PERFORMANCE

Source: clocks.event_reasons.current_limit

The electrical current limit.

Clocks Event Reasons > PSU Alert

clocks_event_reasons.psu_alert

Section: PERFORMANCE

Source: clocks.event_reasons.psu_alert

A power supply alert.

Clocks Event Reasons > SW Range

clocks_event_reasons.sw_range

Section: PERFORMANCE

Source: clocks.event_reasons.sw_range

The frequency range set in software, e.g. with config --frequencyrange.

Clocks Event Reasons > HW Range

clocks_event_reasons.hw_range

Section: PERFORMANCE

Source: clocks.event_reasons.hw_range

The frequency range the hardware allows.

Clocks Event Reasons > Voltage

clocks_event_reasons.voltage

Section: PERFORMANCE

Source: clocks.event_reasons.voltage

The voltage limit.

Clocks Event Reasons > Thermal Alert

clocks_event_reasons.thermal_alert

Section: PERFORMANCE

Source: clocks.event_reasons.thermal_alert

A thermal alert.

Clocks Event Reasons > Power Alert

clocks_event_reasons.power_alert

Section: PERFORMANCE

Source: clocks.event_reasons.power_alert

A power alert.

Clocks Event Reasons Counters > Throttled Time

clocks_event_reasons_counters.throttled_time

Section: PERFORMANCE

Unit: us

Source: clocks.throttle_time

Time the GPU clock has been held back by any throttle reason, accumulated by the driver. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not. The Linux runtime does not count it.

Memory and compute

Row

Details

Meaning

FB Memory Usage > Total

fb_memory_usage.total

Section: MEMORY

Unit: MiB

Source: memory.total

Device memory the driver manages.

FB Memory Usage > Reserved

fb_memory_usage.reserved

Section: MEMORY

Unit: MiB

Source: memory.reserved

Memory the driver keeps back: installed size minus Total. Used includes it, so Total is Used plus Free.

FB Memory Usage > Used

fb_memory_usage.used

Section: MEMORY

Unit: MiB

Source: memory.used

Device memory in use.

FB Memory Usage > Free

fb_memory_usage.free

Section: MEMORY

Unit: MiB

Source: memory.free

Device memory free.

BAR Memory > Total

bar_memory.total

Section: PCI

Unit: MiB

Source: pci.bar.memory.size

Size of the device’s largest memory BAR, the window through which the host reaches its local memory. Nothing reports how much of it is in use.

BAR Memory > Resizable BAR

bar_memory.resizable_bar

Section: PCI

Source: pci.bar.resizable

Whether Resizable BAR is on for that BAR: Enabled or Disabled. The driver reads it from PCI configuration space, which needs CAP_SYS_ADMIN.

Memory Info > Memory Type

memory_info.memory_type

Section: MEMORY

Source: discovery Memory Type

Kind of device memory, e.g. GDDR6 or LPDDR5X.

Memory Info > Number of Memory Channels

memory_info.number_of_memory_channels

Section: MEMORY

Source: discovery Number of Memory Channels

Memory channels. N/A when the driver does not know them.

Memory Info > Memory Bus Width

memory_info.memory_bus_width

Section: MEMORY

Unit: bits

Source: discovery Memory Bus Width

Width of the memory bus. N/A when the driver does not know it.

Scheduler Mode

scheduler_mode

Section: COMPUTE

Source: scheduler.mode

How the device schedules work: timeout, timeslice or exclusive, as set with config.

Compute Resources > Number of Tiles

compute_resources.number_of_tiles

Section: COMPUTE

Source: discovery Number of Tiles

Tiles (sub-devices) on the device.

Compute Resources > Number of Slices

compute_resources.number_of_slices

Section: COMPUTE

Source: discovery Number of Slices

Slices in each tile.

Compute Resources > Number of Sub-slices per Slice

compute_resources.number_of_sub_slices_per_slice

Section: COMPUTE

Source: discovery Number of Sub-slices per Slice

Sub-slices in each slice.

Compute Resources > Number of EUs

compute_resources.number_of_eus

Section: COMPUTE

Source: discovery Number of EUs

Execution units on the device.

Utilization

Measured over the sampling window. Engine utilization needs CAP_PERFMON; without it the values come from the processes’ DRM file statistics where the kernel provides them.

Row

Details

Meaning

Utilization > GPU

utilization.gpu

Section: UTILIZATION

Unit: %

Source: utilization.gpu

Share of the window the busiest engine was busy: from the engine counters, otherwise from the per-process engine times in fdinfo. Never from the EU array, whose metric stream is slow to open; --query-gpu eu.active reports that figure.

Utilization > Memory

utilization.memory

Section: MEMORY, UTILIZATION

Unit: %

Source: utilization.memory

Memory bandwidth used, as a share of the most the memory allows.

Utilization > Compute

utilization.compute

Section: UTILIZATION

Unit: %

Source: utilization.compute

Busiest compute engine. N/A on a device without compute engines.

Utilization > Render

utilization.render

Section: UTILIZATION

Unit: %

Source: utilization.render

Busiest render engine. N/A on a device without render engines, such as a compute-only accelerator.

Utilization > Media

utilization.media

Section: UTILIZATION

Unit: %

Source: utilization.media

Busiest media engine.

Utilization > Video

utilization.video

Section: UTILIZATION

Unit: %

Source: utilization.video

Busiest video decode/encode engine. Linux runs decode and encode on one engine class, so they share this figure. From the engine counters, otherwise fdinfo; never the EU array.

Utilization > Video Enhancement

utilization.video_enhancement

Section: UTILIZATION

Unit: %

Source: utilization.video_enhancement

Busiest video enhancement engine (scaling, color conversion).

Utilization > Copy

utilization.copy

Section: UTILIZATION

Unit: %

Source: utilization.copy

Busiest copy engine.

Temperature

The shutdown and slowdown rows are configured thresholds, the ones the health command compares against, not sensor readings.

Row

Details

Meaning

Temperature > GPU Current Temp

temperature.gpu_current_temp

Section: TEMPERATURE

Unit: C

Source: temperature.gpu

GPU core temperature. N/A on a device that exposes no temperature sensor, or when the driver cannot read the sensor, which on some devices needs root.

Temperature > GPU Shutdown Temp

temperature.gpu_shutdown_temp

Section: TEMPERATURE

Unit: C

Source: temperature.gpu.threshold.shutdown

GPU core temperature at which the device shuts down: the threshold configured for the device, otherwise xpu-smi’s default (130 C) lowered to the driver’s maximum for the sensor (125 C on CRI).

Temperature > GPU Slowdown Temp

temperature.gpu_slowdown_temp

Section: TEMPERATURE

Unit: C

Source: temperature.gpu.threshold.throttle

GPU core temperature at which the clock is throttled.

Temperature > GPU Max Operating Temp

temperature.gpu_max_operating_temp

Section: TEMPERATURE

Unit: C

Source: temperature.gpu.max

Highest temperature the driver says the GPU sensor can operate at; 125 C on CRI. Left out where the driver reports none.

Temperature > Memory Current Temp

temperature.memory_current_temp

Section: TEMPERATURE

Unit: C

Source: temperature.memory

Memory temperature. N/A on a device that exposes no memory sensor, or when the driver cannot read the sensor, which on some devices needs root.

Temperature > Memory Shutdown Temp

temperature.memory_shutdown_temp

Section: TEMPERATURE

Unit: C

Source: temperature.memory.threshold.shutdown

Memory temperature at which the device shuts down: the configured threshold, otherwise xpu-smi’s default (100 C) lowered to the driver’s maximum, if lower.

Temperature > Memory Slowdown Temp

temperature.memory_slowdown_temp

Section: TEMPERATURE

Unit: C

Source: temperature.memory.threshold.throttle

Memory temperature at which the clock is throttled.

Temperature > Memory Max Operating Temp

temperature.memory_max_operating_temp

Section: TEMPERATURE

Unit: C

Source: temperature.memory.max

The same for the memory sensor.

Power and clocks

Row

Details

Meaning

GPU Power Readings > Average Power Draw

gpu_power_readings.average_power_draw

Section: POWER

Unit: W

Source: power.draw

Power the card drew, averaged over the sampling window.

GPU Power Readings > Instantaneous Power Draw

gpu_power_readings.instantaneous_power_draw

Section: POWER

Unit: W

Source: power.draw.instant

Card power at one moment, as the driver reports it (zesPowerGetUsage). Needs a runtime and loader that support it; on CRI it comes from telemetry, which needs root.

GPU Power Readings > GPU Domain Power Draw

gpu_power_readings.gpu_domain_power_draw

Section: POWER

Unit: W

Source: power.draw.gpu

The GPU power domain’s share, averaged over the sampling window.

GPU Power Readings > Current Power Limit

gpu_power_readings.current_power_limit

Section: POWER

Unit: W

Source: power.limit

Power limit in force: the sustained limit, or the burst or peak limit on a device without one.

GPU Power Readings > Burst Power Limit

gpu_power_readings.burst_power_limit

Section: POWER

Unit: W

Source: power.limit.burst

Burst power limit (PL2), on devices that report one.

GPU Power Readings > Peak Power Limit

gpu_power_readings.peak_power_limit

Section: POWER

Unit: W

Source: power.limit.peak

Peak power limit (PL4), on devices that report one. Current Power Limit shows it on a device with no sustained or burst limit.

GPU Power Readings > Default Power Limit

gpu_power_readings.default_power_limit

Section: POWER

Unit: W

Source: power.limit.default

Default power limit from the device properties. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not. CRI’s driver reports it as unknown, so --verbose shows N/A.

GPU Power Readings > Min Power Limit

gpu_power_readings.min_power_limit

Section: POWER

Unit: W

Source: power.limit.min

Lowest power limit that can be set. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not.

GPU Power Readings > Max Power Limit

gpu_power_readings.max_power_limit

Section: POWER

Unit: W

Source: power.max_limit

Highest power limit that can be set. Printed where the driver supplies it; --verbose prints Not Supported where it says the device does not.

GPU Power Readings > Memory Power Draw

gpu_power_readings.memory_power_draw

Section: POWER

Unit: W

Source: power.draw.memory.instant

Memory power domain at one moment, on devices with one (the B580 has one; CRI does not).

Clocks > Graphics

clocks.graphics

Section: CLOCK

Unit: MHz

Source: clocks.current.graphics

Current GPU core clock.

Clocks > Memory

clocks.memory

Section: CLOCK

Unit: MHz

Source: clocks.current.memory

Current memory clock. N/A when the driver does not know it, including a reading of 0 MHz, as on CRI.

Clocks > Media

clocks.media

Section: CLOCK

Unit: MHz

Source: clocks.current.media

Current media engine clock; can read 0 while the media engine is idle.

Applications Clocks > Graphics

applications_clocks.graphics

Section: CLOCK

Unit: MHz

Source: clocks.requested.graphics

GPU core clock the driver last requested, which the GPU runs at unless a clock event reason holds it back.

Max Clocks > Graphics

max_clocks.graphics

Section: CLOCK

Unit: MHz

Source: clocks.max.graphics

Highest GPU core clock the device supports.

Max Clocks > Media

max_clocks.media

Section: CLOCK

Unit: MHz

Source: clocks.max.media

Highest media engine clock the device supports.

Supported Clocks > Graphics

supported_clocks.graphics

Section: SUPPORTED_CLOCKS

Unit: MHz

Source: clocks.supported.graphics

GPU core clocks the minimum and maximum of config --frequencyrange can be set to, as low-high step N when evenly spaced. Printed only when SUPPORTED_CLOCKS or ALL is selected.

Processes

One record per process using the device; None when there is none. In JSON, processes is an array of objects with the keys below.

Row

Details

Meaning

Processes > Process ID

processes[].process_id

Section: PIDS

Source: ps

Process ID.

Processes > Type

processes[].type

Section: PIDS

Source: ps

Engine the process uses: C compute, G render or 3D, M media, DMA copy, Other; C for a process holding memory without naming an engine.

Processes > Name

processes[].name

Section: PIDS

Source: ps

Command name.

Processes > Used GPU Memory

processes[].used_gpu_memory

Section: PIDS

Unit: MiB

Source: ps

Device memory the process holds.

Reading the Report

  • A value the device, driver or platform cannot supply prints as N/A. Where the value needs something the caller lacks, the text report says what, for example N/A (requires CAP_SYS_ADMIN): CAP_SYS_ADMIN for the PCIe link generation and width, the atomic capabilities and Resizable BAR, which come from PCI configuration space; CAP_PERFMON for engine utilization; read-write access to the device’s /dev/mei* node for the GFX firmware version. The report prints no diagnostic for such a value; set XPU_SMI_LOG_LEVEL=DBG to see why one is missing.

  • Measured values (power draw, utilization, throughput, bandwidth) are averaged over a 200 ms window that the first report waits for. In a loop each later report averages over the time since the previous one.

  • Values with a unit are shown to two decimals. JSON keeps the full precision.

  • GPU IP Version is the graphics IP version the driver reports, in hex. It identifies the architecture: newer devices have higher values.

  • The report has no ECC section for now: ECC mode, the ECC and RAS error counts, Offline Pages and Memory Repair are left out until reading the driver’s RAS counters is reliable. Use --query-gpu (for example --query-gpu ecc.errors.aggregate.total) or stats -r for them meanwhile.

  • PCI: Host Max is the link capability of the root port the device is under, the CPU end of the path; a switch in between can limit the link further. PCIe Errors are the device’s AER counters since it was enumerated, and Atomic Caps Inbound and Atomic Requests come from its PCI Express capability, which needs root. All of these are Linux only, and left out where the device or OS does not report them.

  • Fan Speed, the Performance Factors, Throttled Time, PCIe Replays Since Reset, the default, minimum and maximum power limits and Memory Power Draw come from Level Zero calls that many drivers do not implement; CRI’s driver supplies none of them. They are printed where the driver supplies a value, and with --verbose otherwise.

  • GPU Power Readings: Average Power Draw is the card’s draw over the window, Instantaneous Power Draw the driver’s reading at one moment, and GPU Domain Power Draw the GPU domain’s share. Current Power Limit is the sustained limit, or the burst or peak limit on a device without one; the burst and peak limits are listed on devices that report them.

  • BAR Memory is the device’s largest memory BAR, through which the host reaches device memory. Resizable BAR says whether it was resized; the driver reads that from PCI configuration space.

  • FB Memory Usage: Reserved is memory the driver keeps back (installed size minus Total). Used counts it, so Total is Used plus Free. A memory clock the driver reports as 0 reads N/A.

  • Utilization comes from the Sysman engine counters, or from the per-process engine times in fdinfo where those are unavailable; never from the EU array, whose metric stream is slow to open. Video and Video Enhancement split Media into its decode/encode engines (which Linux reports as one) and its enhancement engines, on devices that have them.

  • GPU utilization is the busiest engine’s share of the window. An engine counts as busy for all the time it has work scheduled, so it can read higher than eu.active under the same workload.

  • Applications Clocks shows the graphics clock the driver last requested, which the GPU runs at unless a throttle reason holds it back. Minor Number is the number of the device’s DRM card node, and reads N/A on Windows. Slot Number is the physical PCI slot the platform firmware names, and reads N/A on systems whose firmware names none.

  • GPU Recovery Action says what the driver reports the device needs: None, Reset (for example a wedged device), or Reset to repair.

  • As root, the GFX firmware version is read from the GSC firmware interface, which takes about 200 ms per device on some platforms. Devices are read in parallel, and the read overlaps the measurement window. On a device with an AMC the serial number and the board part number are read from the AMC, one after the other.

  • The slowdown and shutdown temperatures are the configured thresholds the Health command compares against, not sensor readings. Without a threshold configured for the device, the default shutdown temperature is lowered to the driver’s maximum for that sensor, as on CRI (125 C). The max operating temperatures are what the driver reports for each sensor, on devices that report one.

  • On Windows the kernel-mode driver version reads N/A.

  • Rows only some devices or drivers supply, such as the burst and peak power limits, are left out on a device without them, so it does not fill with N/A; the JSON output leaves out the same rows. --verbose prints them all, and Not Supported for a value the driver or OS says the device does not support; a value missing for another or an unknown reason stays N/A.

A device in survivability mode is listed with only its state and the recovery action (Firmware update and GPU reset).

JSON Output

--json prints one object. Its keys are the lower-case, underscore-separated form of the labels in the text report, and it has the same rows: a row the text leaves out is not a key, and --verbose adds them to both. A value that is not available, printed N/A in the text, is null. Numbers are strings in the unit shown in the text report, and processes is an array.

{
    "timestamp": "2026-10-21T13:28:16Z",
    "umd_version": "1.17.39980",
    "kmd_version": null,
    "level_zero_version": "1.33.1",
    "attached_gpus": 1,
    "gpus": [
        {
            "pci_bus_id": "0000:03:00.0",
            "product_name": "Intel(R) Arc(TM) B580 Graphics",
            "fb_memory_usage": {
                "total": "12216",
                "reserved": "0",
                "used": "218.62109375",
                "free": "11997.37890625"
            },
            "processes": []
        }
    ]
}

Examples

Report on every device:

xpu-smi query

Report only memory and power for device 0:

xpu-smi query --device 0 --display MEMORY,POWER

Save a report to a file:

xpu-smi query --file report.txt

Print the power and temperature sections every 5 seconds, 12 times, as JSON:

xpu-smi query --display POWER,TEMPERATURE --json --loop 5 --count 12

See Also

  • Discovery for static device properties

  • Stats for statistics collected over an interval

  • Dump for metrics in columns, one row per sample