Dump

Dump device statistics data continuously at a configurable interval. Suitable for time-series data collection and high-frequency sampling.

Synopsis

xpu-smi dump -d [deviceIds] --metrics [metricsSpec] --interval [seconds] --number [count]
xpu-smi dump --device [deviceIds] --metrics [metricsSpec] --interval [seconds] --number [count]
xpu-smi dump --device [deviceIds] --metrics [metricsSpec] --file [filename] --loop-ms [milliseconds] --time [seconds]

Options

-h, --help

Print this help message and exit.

-j, --json

Print result in JSON format.

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

Device IDs or PCI BDF addresses to query. Use -1 to select all devices. Multiple IDs can be separated by a comma.

--metrics <metricsSpec>, --select <metricsSpec>

Metrics to collect. Accepts any of the following forms, or a comma-separated mix:

  • Legacy numeric IDs: 0, 1, 0,1,2 (see Metrics Reference below)

  • Group names: POWER, UTILIZATION, TEMPERATURE, MEMORY, CLOCK, PCI, ECC, EU_ARRAY, FAN, ALL

  • Single-char shortcuts: p (Power + Temperature), u (Utilization), m (Memory), c (Clock), t (PCI), e (ECC), x (EU_ARRAY), f (Fan)

  • Multi-char combos: pu (Power + Temperature + Utilization), pum (Power + Temperature + Utilization + Memory)

Note

Dot-notation field names (e.g. temperature.gpu) are not accepted by --metrics. Use --query-gpu at the top level for field-name queries.

--interval <seconds>, --delay <seconds>, --loop <seconds>

Sampling interval in seconds between dumps. Default: 1. Maximum: 20. --delay and --loop are aliases for --interval.

--number <count>, --count <count>

Number of samples to dump. If omitted, dumps continuously until the user terminates the operation (Ctrl-C).

--loop-ms <milliseconds>

Millisecond-precision sampling interval. Overrides --interval when present. Use for sub-second sampling. Recommend pairing with --file.

Note

A warning is emitted when --loop-ms is below 50 ms and power metrics are selected, as the power measurement window may be too short for accurate readings.

-f <filename>, --file <filename>, --filename <filename>

Write output to a file instead of stdout.

--time <seconds>

Total dump duration in seconds. Cannot be combined with --number.

--date

Include date (YYYY/MM/DD) prefix in each timestamp column.

--format <spec>

Output format for CSV output. Comma-separated flags:

  • csv — CSV output (default when --file is used)

  • noheader — suppress the header row

  • nounits — omit unit suffixes (e.g. (W), (C)) from headers

Example: --format csv,noheader,nounits

Metrics Reference

ID

Metric

0

GPU Utilization (%), per tile or device. Device-level is the average value of tiles for multi-tile devices.

1

GPU Power (W), per tile or device

2

GPU Frequency (MHz), per tile or device

3

GPU Core Temperature (Celsius), per tile or device

4

GPU Memory Temperature (Celsius), per tile or device

5

GPU Memory Utilization (%), per tile or device

6

GPU Memory Read (kB/s), per tile or device

7

GPU Memory Write (kB/s), per tile or device

8

GPU Energy Consumed (J), per tile or device

9

GPU EU Array Active (%), per tile or device. At least one thread is active. Device-level is the average value of tiles for multi-tile devices.

10

GPU EU Array Stall (%), per tile or device. At least one thread is loaded but the EU is stalled. Device-level is the average value of tiles for multi-tile devices.

11

GPU EU Array Idle (%), per tile or device. Device-level is the average value of tiles for multi-tile devices.

12

Reset Counter

13

Programming Errors

14

Driver Errors

15

Cache Errors Correctable

16

Cache Errors Uncorrectable

17

GPU Memory Bandwidth Utilization (%), per tile or device. Device-level is the average value of tiles for multi-tile devices.

18

GPU Memory Used (MiB), per tile or device

19

PCIe Read (kB/s), per device

20

PCIe Write (kB/s), per device

21

Unsupported

22

Compute Engine Utilization (%), per tile

23

Render Engine Utilization (%), per tile

24

Media Decoder Engine Utilization (%), per tile

25

Media Encoder Engine Utilization (%), per tile

26

Copy Engine Utilization (%), per tile

27

Unsupported

28

Unsupported

29

GPU Memory Errors Correctable. Device-level is the sum value of tiles for multi-tile devices.

30

GPU Memory Errors Uncorrectable. Device-level is the sum value of tiles for multi-tile devices.

31

Compute Engine Group Utilization (%), per tile

32

Render Engine Group Utilization (%), per tile

33

Media Engine Group Utilization (%), per tile

34

Copy Engine Group Utilization (%), per tile

35

Throttle reason, per tile

36

Media Engine Frequency (MHz), per tile or device

Examples

Dump GPU utilization and power for all devices every second, 10 times:

xpu-smi dump --device -1 --metrics 0,1 --interval 1 --number 10

Dump using group name shortcuts:

xpu-smi dump --device 0 --metrics pu --interval 1

Dump frequency and temperature for device 0 every 2 seconds:

xpu-smi dump --device 0 --metrics CLOCK,TEMPERATURE --interval 2

Millisecond-precision dump to file:

xpu-smi dump --device 0 --metrics POWER,UTILIZATION --file metrics.csv --loop-ms 100 --time 60

Dump with date in timestamp, CSV with no header or units:

xpu-smi dump --device 0 --metrics 0,1 --date --format csv,noheader,nounits