Targeted Power Plugin#
The Targeted Power plugin detects a GPU that cannot reach a configured minimum board-power threshold under sustained compute load. Such a result can expose a power-delivery, GPU loading, or performance limitation; an enforced power limit that makes the target unattainable is reported separately as a configuration constraint.
Availability and prerequisites#
The canonical test name is targeted_power. It runs on selected physical
GPUs. Run it on idle GPUs: competing workloads and an enforced power limit below
the configured target can prevent a meaningful target-power result.
Behavior#
This test’s core purpose is to sustain a high level of power usage. It relies on CUDA and performs large matrix multiplications simultaneously on each GPU in order to keep the GPUs busy and drawing power. Each GPU has a large workload that is sustained throughout the test; the workload does not pulse.
Parameters#
The following table lists the global parameters for the targeted power plugin:
Parameter |
Type |
Default |
Description |
|---|---|---|---|
test_duration |
Double |
120.0 |
This is the time in seconds that the test should run. |
target_power |
Double |
Product-specific; plugin fallback is 100.0 |
Target board power, in watts. The installed SKU configuration normally overrides the plugin fallback. |
target_power_min_ratio |
Double |
0.75 |
Minimum fraction of |
use_dgemm |
Bool |
True |
If set to true, the test will use 64 bit precision in its matrix multiplications instead of 32 bit. |
is_allowed |
Bool |
False |
Specifies whether or not this test is allowed to run. |
Examples#
Run the power test for 10 minutes:
$ dcgmi diag -r targeted_power -p targeted_power.test_duration=600.0
Run the level 3 diagnostic with a 5 minute targeted power test:
$ dcgmi diag -r 3 -p targeted_power.test_duration=300.0
Run the target power test targeting 200 W of power usage:
$ dcgmi diag -r targeted_power -p targeted_power.target_power=200.0
Run the level 4 test, skipping targeted power:
$ dcgmi diag -r 4 -p targeted_power.is_allowed=false
Run the targeted power test, using single precision (32 bit):
$ dcgmi diag -r targeted_power -p targeted_power.use_dgemm=false
Results and failure conditions#
Condition |
Result or code |
Interpretation |
|---|---|---|
Maximum observed board power is below
|
|
The GPU did not reach the configured power threshold. The message can include a utilization note when another workload affected the run. |
The enforced power limit is too low for the configured minimum to be attainable. |
|
This is a configuration constraint, not evidence that the GPU failed under an attainable target. |
DCGM cannot obtain the board-power statistic needed for evaluation. |
|
The test cannot make its target-power determination. |
A common monitored condition occurs, including an uncorrectable memory error, XID, failed or pending row remap, thermal violation, or an enabled SBE threshold violation. |
Fail with the condition-specific |
See Diagnostic Errors for the reported condition. |
CUDA setup, host-memory allocation, a DCGM call, or another execution operation fails. |
|
The workload did not complete. These are not target-power findings. |
The plugin is disabled. |
|
Numbered-suite or product configuration did not allow the workload. |
The run is interrupted. |
|
No completed target-power result was produced. |