Pulse Test Diagnostic#

The Pulse Test evaluates whether the board and platform power-delivery system can tolerate rapid changes in GPU power and current demand. It creates load transitions intended to expose power-delivery instability. A successful run means that the selected workload did not expose such instability; it is not a comprehensive certification of the power supply.

Availability and prerequisites#

The canonical test name is pulse_test. It runs on selected physical GPUs. The required pulse workload and its supported pattern set are product- and release-specific. Run the plugin on an idle system because it deliberately produces rapid power and current changes.

Behavior#

The plugin launches the supplied pulse workload for each selected pattern and records that workload’s completion status and reported errors. The aggregate result caveat is described below. The product-specific default patterns have been verified with oscilloscopes to produce worst-case power transients.

Parameters#

Pulse Test parameters#
Parameter	Type	Default	Description
`is_allowed`	Boolean	`false`	Allow the plugin to run. `false` produces a skipped result.
`test_duration`	Number	`60`	Seconds requested for each selected pattern. Total runtime also includes setup and all selected patterns.
`patterns`	String	All	Comma-separated pattern indices. Available indices are product- and release-specific; an index does not constitute a stable description of the underlying workload.

If patterns contains both valid and unsupported indices, DCGM ignores the unsupported indices and runs the valid subset. A nonnumeric value or a list with no valid index produces DCGM_FR_BAD_PARAMETER. Use an explicit pattern list only when directed by documentation for the installed release or by NVIDIA support.

Examples#

Run the entire diagnostic suite, including the pulse test:

$ dcgmi diag -r 4

Run just the pulse test:

$ dcgmi diag -r pulse_test

Run only selected pulse patterns:

$ dcgmi diag -r pulse_test -p pulse_test.patterns=0,1,2

Run the selected patterns with a 180-second iteration duration:

$ dcgmi diag -r pulse_test -p "pulse_test.patterns=0,1,2;pulse_test.test_duration=180"

Results and failure conditions#

Pulse Test results#
Condition	Result or code	Interpretation
Every selected pulse workload reports successful completion.	Pass	No instability was observed by this workload. This is not a general certification of the platform power supply.
The workload reports an error or an unsuccessful completion status.	`DCGM_FR_INTERNAL`; remaining patterns stop	Inspect the accompanying workload error and diagnostic log. The symptom can be consistent with power-delivery instability but does not identify its cause by itself. In DCGM 4.6 the aggregate result can remain pass or skip, so do not discard the reported error merely from the result label.
The workload exits nonzero, is terminated by a signal, or returns no usable completion status.	`DCGM_FR_INTERNAL`; remaining patterns stop	The pulse run did not complete normally; distinguish execution and environment problems from a repeatable under-load failure. Inspect the error even if the aggregate result is not labeled fail.
The required workload executable cannot be found or launched, or its output cannot be read.	`DCGM_FR_INTERNAL`; fail	This is a setup or execution failure and makes no statement about GPU or power-delivery health.
Pattern syntax is invalid or none of the requested indices is valid for the selected product.	`DCGM_FR_BAD_PARAMETER`; no usable pattern result	Correct or omit `patterns`. Mixed lists run only their valid subset; in DCGM 4.6 a wholly invalid list can leave the aggregate result skipped rather than failed.
DCGM cannot retrieve device information needed to select the supported workload.	`DCGM_FR_DCGM_API`; fail	The pulse workload was not started.
`is_allowed=false`.	`DCGM_FR_TEST_DISABLED`; skip	No pulse workload was run and no health result was produced.