dcgmi config#

NAME#

dcgmi-config - get, set, and enforce GPU configuration

SYNOPSIS#

dcgmi config --get [--group <group-id>] [--verbose] [--json]
             [--host <address>]
dcgmi config --set [--group <group-id>] <configuration-options>
             [--host <address>]
dcgmi config --enforce [--group <group-id>] [--host <address>]

DESCRIPTION#

dcgmi config reads and changes the target configuration maintained by DCGM for a GPU group. --set updates the requested properties and applies them. --enforce reapplies the stored target configuration when current GPU state has drifted. The default target is the built-in all-GPU group.

Changing GPU configuration generally requires administrative privileges and can affect running workloads.

OPTIONS#

Operation selection#

--get: Display current and target configuration. The report includes supported properties such as application clocks, ECC mode, power limit, compute mode, and workload power profiles.
--set: Set one or more configuration properties. At least one configuration option is required.
--enforce: Reapply the target configuration to every GPU in the selected group.

Targets and output#

-g <group-id>, --group <group-id>: Select a persistent GPU group. The default is the built-in all-GPU group.
-v, --verbose: With --get, display configuration separately for each GPU.
-j, --json: With --get, format the configuration body as JSON. DCGM 4.6 can append a plain-text homogeneity notice after the JSON document. See JSON output.

Configuration options#

-e <0|1>, --eccmode <0|1>: Disable or enable ECC mode.
-s <0|1>, --syncboost <0|1>: Disable or enable synchronized boost for the group.
-a <memory,sm>, --appclocks <memory,sm>: Set the memory and SM application clocks in MHz.
-P <watts>, --powerlimit <watts>: Set the per-GPU power limit in watts. The value must be nonnegative and within the limits supported by each GPU.
-c <mode>, --compmode <mode>: Set the compute mode: 0 for unrestricted, 1 for prohibited, or 2 for exclusive-process mode.
-o <index>, --workloadpowerprofile <index>: Select a device-specific workload power-profile index. Use -1 to clear every requested profile. Nonnegative indices use the action selected by --workloadpowerprofileaction.
-w <action>, --workloadpowerprofileaction <action>: Choose how --workloadpowerprofile changes the requested profiles: a adds the selected profile, c clears it, and o overwrites the current selection. The default is a.

Common options#

--host <address>: Connect to the specified host engine.
-h, --help: Print generated usage information and exit.

CONFIGURATION VALUE REFERENCE#

Configuration support is GPU-specific. Before configuring a heterogeneous group, query each member with config --get --verbose and the discovery command in the following table.

Configuration properties and valid values#
Property	Accepted value	How to determine support
ECC mode	`0` disabled; `1` enabled	`dcgmi config --get --verbose` reports `Not Supported` where the property is unavailable
Synchronized boost	`0` disabled; `1` enabled	`dcgmi config --get --verbose` reports current support and state; every GPU in the group must be compatible with the group setting
Application clocks	An exact `<memory-MHz>,<SM-MHz>` pair	`dcgmi discovery --info c --gpuid <id>` lists every supported pair, in memory-clock then SM-clock order
Power limit	A nonnegative integer number of watts within the device range	`dcgmi discovery --info p --gpuid <id>` lists the minimum and maximum and also shows current, default, and enforced limits
Compute mode	`0` unrestricted; `1` prohibited; `2` exclusive process	`dcgmi config --get --verbose` reports current support and state
Workload power profile	A supported nonnegative profile index; `-1` clears all requested profiles. Actions are `a` add, `c` clear, and `o` overwrite.	`dcgmi discovery --info w --gpuid <id>` lists supported, requested, and enforced indices. `nvidia-smi power-profiles -l` supplies profile names on supported systems; see Workload Power Profile Configuration.

If a discovery query reports that a feature is unsupported, there is no valid value for that property on that GPU. For a group, use a clock pair, power limit, or profile index that is valid for every member.

APPLICATION AND LIFETIME#

--set applies every requested property to every GPU in the group and records the target configuration in the host engine. If any property fails on any GPU, the command returns an error and identifies the failed GPU/property. The operation is not a transaction: do not assume that earlier successful changes were rolled back after a later failure. Verify the result with config --get --verbose.

Target configuration survives the dcgmi client disconnecting and is automatically reapplied after a GPU reset or reinitialization. It is host-engine state, not durable on-disk configuration, so it is not restored after the host-engine process or system restarts. The applied settings belong to the individual GPUs and remain when the group used to set them is deleted. --enforce requires an existing target configuration and manually reapplies it at any time.

The current state can differ from the target because a property is pending, unsupported, or was changed outside DCGM. Not Specified in the target column means DCGM has no target for that property; it is not a request to restore a hardware default.

EXAMPLES#

Display current and target configuration for all GPUs:

$ dcgmi config --get

Set group 1 to exclusive-process mode with a 300 W power limit:

$ dcgmi config --set --group 1 --compmode 2 --powerlimit 300

Add workload power profile 2 to group 5:

$ dcgmi config --set --group 5 --workloadpowerprofile 2 \
    --workloadpowerprofileaction a

Clear all requested workload power profiles:

$ dcgmi config --set --workloadpowerprofile -1

Reapply the target configuration for group 1:

$ dcgmi config --enforce --group 1