RAS Log

Collect GPU RAS (Reliability, Availability, Serviceability) hardware error records. By default the read is non-destructive (peek): records remain in the hardware error buffer and can be read again by subsequent calls. Use --drain to consume records from the buffer; a --file destination is required in that mode to avoid data loss.

Note

This command is supported on Linux only.

Synopsis

xpu-smi raslog [-t cper]
xpu-smi raslog [-t cper] -j
xpu-smi raslog [-t cper] -f [fileName]
xpu-smi raslog [-t cper] -f [fileName] --drain
xpu-smi raslog [-t cper] [-f fileName] --instance [name]
xpu-smi raslog [-t cper] [-f fileName] --buffer-size-kb [kb]

Options

-h, --help

Print this help message and exit.

-j, --json

Print status in JSON format. The raw CPER binary is still written to the file specified by --file; the JSON output describes the collection result and per-record metadata only.

-t <type>, --type <type>

The hardware log type to collect. Currently the only supported type is cper (Common Platform Error Record), which is also the default.

-f <fileName>, --file <fileName>

The file to write the raw hardware log data into. Optional in the default peek mode; required with --drain since consumed records cannot be recovered. When provided, the path is validated before any data is read.

--drain

Read and consume records from the hardware error buffer (destructive mode). Records will not be returned by subsequent reads. --file is required in this mode to ensure no records are lost.

--instance <name>

Collect from the named hardware trace instance instead of the global error buffer. Requires driver support; an error is returned if named instances are not available on the current system. Each invocation creates a fresh collection instance; if an instance with the same name is already active, the command fails with a HANDLE_OBJECT_IN_USE error.

--buffer-size-kb <kb>

Request a specific total error buffer size in kilobytes across all per-CPU buffers. The driver splits and rounds the value as needed. If omitted, the driver default size is used.

Output

Without --json, a single status line is printed. The exact wording depends on whether --file and --drain were supplied:

# peek (default), no --file
CPER buffer (4096 bytes, 3 records) collected (records not consumed).

# peek with --file
CPER buffer (4096 bytes, 3 records) copied (records not consumed) to file: cper.bin

# --drain with --file
CPER buffer (4096 bytes, 3 records) written to file: cper.bin

With --json, a JSON object is printed to stdout:

{
    "status": "OK",
    "bytes": 4096,
    "buffer_bytes": 4096,
    "records": 3,
    "file": "cper.bin",
    "cper_records": [
        {
            "offset": 0,
            "length": 1400,
            "bdf": "0000:4d:00.0",
            "uuid": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
            "timestamp_ns": 1749123456789012,
            "record_type": "ERROR_CORRECTED"
        }
    ]
}

JSON fields

Field	Description
status	"OK" on a complete read; "PARTIAL" if some records were too large to fit in the buffer and were dropped.
bytes	Bytes written to the output file. Zero when --file is not provided.
buffer_bytes	Bytes collected from the hardware error buffer, regardless of whether a file was written. Equal to bytes when a file was written.
records	Number of error records collected.
file	Path of the output file. Present only when --file was provided.
warning	Present only when status is "PARTIAL". Describes the reason records were dropped.
cper_records	Array of per-record metadata objects (see below).

Per-record metadata (cper_records entries)

Field	Description
offset	Byte offset of this record within the collected CPER blob. When --file is provided this matches the byte offset within the output file.
length	Length of this record in bytes.
bdf	PCI Bus/Device/Function address of the GPU that reported the error, in domain:bus:device.function format.
uuid	UUID identifying the GPU device that reported the error.
timestamp_ns	Timestamp of the error event in nanoseconds. The reference point is implementation-defined and only consistent across records of the same info log instance.
record_type	Severity of the record: UNKNOWN, INFORMATIONAL, ERROR_CORRECTED, ERROR_RECOVERABLE, or ERROR_FATAL.

Error output

On failure, a JSON error object is printed to stdout when --json is active:

{
    "ze_result": 2013265923,
    "error": "CPER hardware log is not supported on this system or build"
}

Error JSON fields

Field	Description
ze_result	The ze_result_t error code as a decimal integer. Common values: 2013265923 (ZE_RESULT_ERROR_UNSUPPORTED_FEATURE), 2147483646 (ZE_RESULT_ERROR_UNKNOWN).
error	Human-readable description of the failure.
file	Present only for file I/O errors; path of the output file that could not be opened or written.

Examples

Peek at hardware error records (non-destructive, no file required):

xpu-smi raslog

Peek and print a JSON summary:

xpu-smi raslog -j

Peek and save the raw binary to a file:

xpu-smi raslog -f cper.bin

Drain (consume) records into a file:

xpu-smi raslog --drain -f cper.bin

Drain and print a JSON summary:

xpu-smi raslog --drain -f cper.bin -j

Collect from a named trace instance:

xpu-smi raslog -f cper.bin --instance xpu-collection

Collect with a specific buffer size:

xpu-smi raslog -f cper.bin --buffer-size-kb 8192