Power and Thermal Profiles Co-working Guide#

Overview#

Use the power and thermal profiles together to control both sides of the same thermal budget:

  • set_power_profile.sh limits heat generation by setting a package-power envelope (RAPL PL1/PL2 and intel_lpmd tuning).

  • set_thermal_profile.sh sets the response when package temperature rises (fan, processor frequency cap, then intel_powerclamp).

The power profile is the upper bound on sustained package power. The thermal profile is a protection policy, not a second power-profile selector. A thermal trip can reduce performance below the power target, but it cannot raise power above the RAPL cap.

Run both tools directly on the target system. On a provisioned node their absolute paths are under /opt/edge/developer/tools/power-tuning/.

Acronyms and Terms#

Term

Meaning

cTDP

Configurable Thermal Design Power. The platform-defined range of sustained processor power; Level 2 is the highest supported target.

intel_lpmd

Intel Low Power Mode daemon. The power-profile tool configures it to apply CPU energy-performance tuning.

MSR

Model-Specific Register. Low-level CPU register used to program the RAPL power limits.

PL1

Power Limit 1. The long-term, sustained package-power limit.

PL2

Power Limit 2. The short-duration burst package-power limit above PL1.

PPCC

Power Participant Capabilities Collection. The thermald configuration block that supplies a package power limit to thermald.

PkgTmp

CPU package temperature reported by turbostat, in degrees Celsius.

PkgWatt

CPU package power reported by turbostat; includes CPU cores, integrated GPU, and uncore power.

RAPL

Running Average Power Limit. Intel hardware interface used to read and enforce package and platform power limits.

SysWatt

Whole-platform power reported by turbostat through the optional RAPL psys domain.

psys

RAPL platform-power domain that supplies the SysWatt limit and reading when supported by the system firmware.

Interaction and Dependency#

The profiles intentionally share the package RAPL PL1 setting. When applying a thermal profile, the script reads the current package PL1 and writes it into the generated thermald configuration as a PPCC maximum. thermald then applies that value when it starts.

This makes the order mandatory:

  1. Apply the power profile.

  2. Apply the thermal profile.

  3. Start monitoring before running a workload or stress test.

If thermal is applied first, or thermald is restarted after a later power change, thermald can reset the package PL1 to the PPCC value saved in its existing configuration. Applying the thermal profile again after the power profile refreshes that PPCC value. Always use the same order when changing the power target.

On platforms without the RAPL psys domain, SysWatt is not supported. Track PkgWatt instead as the effective power measurement while validating the combined profiles.

Combination

Expected behavior

Operational guidance

Low power + cool

Lowest temperatures; passive trips may reduce throughput early.

Suitable for fanless or tightly enclosed systems.

Balanced power + warm

Normal balance of throughput, acoustics, and temperature.

Good starting point for qualification.

High power + hot or thermal-max

More performance until higher temperatures; greater throttle and enclosure risk.

Validate under the real workload before deployment.

Any power + thermal trip reached

Processor or intel_powerclamp reduces performance to hold temperature.

Lower the power target or choose earlier trips when sustained.

Supported Profile Combinations#

All named profiles can be combined. Each entry gives the operating posture, not a guarantee that the enclosure can sustain the selected power target. Validate high-power combinations under the intended workload.

Power profile

cool (55/70/80 C)

warm (60/75/85 C)

hot (70/90/95 C)

thermal-max (95/100/104 C)

LowPower (10 W)

Maximum thermal margin.

Balanced low-power operation.

Normally little thermal intervention.

Least protective and seldom needed.

BalancedLow (15 W)

Favors temperature over throughput.

General low-power default.

Permits higher package temperature.

Validate the enclosure before use.

BalancedHigh (20 W)

May reach passive trips under load.

Recommended starting point.

Higher temperature and throttle risk.

Qualification required.

Performance (25 W)

Thermal policy can reduce sustained throughput.

Validate sustained workloads.

Increased heat and fan demand.

High-risk for constrained enclosures.

MaxPerformance (cTDP Level 2)

Thermal trips commonly limit effective throughput.

Require enclosure qualification.

High thermal and throttle risk.

Use only after thermal qualification.

Custom

Valid explicit targets and trips; apply power first, then regenerate thermal.

Valid explicit targets and trips; apply power first, then regenerate thermal.

Valid explicit targets and trips; apply power first, then regenerate thermal.

Valid explicit targets and trips; apply power first, then regenerate thermal.

Conflict and Override Policies#

Condition

Effective policy

Required action

Power profile is applied before thermal profile

The thermal PPCC maximum is refreshed with the current RAPL PL1.

This is the required order.

Thermal profile is applied before, or restarted after, a power-profile change

thermald can restore the older PPCC package cap, overriding the newly intended RAPL PL1.

Re-apply the power profile, then re-apply the thermal profile.

A Processor or intel_powerclamp trip is active

The thermal policy takes precedence for safety and reduces frequency or injects idle cycles; effective power and performance can fall below PL1.

Lower the power profile or select earlier trips if sustained throttling is unacceptable.

Requested PL1 or PL2 exceeds the firmware cTDP limit

Firmware clamps the RAPL request; the lower enforced limit wins.

Adjust BIOS cTDP configuration or use the reported effective limit.

--sysWatt is supplied but psys is unsupported

The SysWatt request is ignored; only PkgWatt is capped.

Track PkgWatt and plan to the package-power limit.

SysWatt reads 0.00 on a psys-capable platform

The psys counter may be frozen even though a cap was written.

Use PkgWatt as the observable power measurement.

OS restart

RAPL PL1/PL2 reset to firmware defaults while thermald PPCC remains on disk.

Re-apply power first, then thermal, before the next workload.

Prerequisites#

  • Intel x86_64 host, msr-tools, the msr kernel module, and thermald.

  • BIOS permits OS control of package power limits and has the package power limit MSR unlocked. See the Power Profiles User Guide.

  • Fan, Processor, or intel_powerclamp cooling devices are available:

    grep . /sys/class/thermal/cooling_device*/type

  • Before the first change, preserve any local configuration that is not already backed up. The thermal tool’s .bak files are replaced on each later apply.

    sudo cp -a /etc/thermald/thermal-conf.xml /etc/thermald/thermal-conf.xml.before-power-thermal
sudo cp -a /etc/systemd/system/thermald.service.d/override.conf /etc/systemd/system/thermald.service.d/override.conf.before-power-thermal

    If either source file does not exist, record that fact instead of creating an empty replacement. The power tool creates a one-time .orig backup only for model-specific intel_lpmd configuration files that it replaces.

Enable Both Profiles#

Preview each profile first. The following uses a 20 W power ceiling and the warm thermal policy:

tools/power-tuning/set_power_profile.sh --profile BalancedHigh --dry-run
tools/power-tuning/set_thermal_profile.sh --profile warm --dry-run

Apply in the required order:

sudo tools/power-tuning/set_power_profile.sh --profile BalancedHigh
sudo tools/power-tuning/set_thermal_profile.sh --profile warm

For a custom package target, apply it before generating the thermal policy so the PPCC maximum captures the intended PL1:

sudo tools/power-tuning/set_power_profile.sh --profile Custom --pkgWatt 30 --sysWatt 35 --burstRatio 1.2
sudo tools/power-tuning/set_thermal_profile.sh --profile custom --fan 60 --proc 75 --clamp 85

Confirm the effective state:

systemctl is-active intel_lpmd.service
systemctl is-active thermald
systemctl show thermald -p ExecStart
cat /sys/class/powercap/intel-rapl/intel-rapl:0/constraint_0_power_limit_uw

The thermald command must contain --ignore-default-control and must not contain --adaptive. The RAPL value is in microwatts; compare it with the requested PL1 after allowing for any firmware cTDP clamp reported by the power script.

Validate Under Load#

Start the monitor before the workload so it captures the ramp, then run a bounded test in a second terminal. Stop the monitor with Ctrl-C after the stress command completes:

sudo tools/power-tuning/pt_mon.sh
sudo tools/power-tuning/stress_gen.sh --gpu 12 --duration 3m

Note: Use stress_gen.sh, openvino_stress.sh, an magic9 benchmark, or your production workload to generate load for power and thermal profiling. Choose a bounded run and use the same workload when comparing profiles.

Watch PkgTmp against the thermal trip points and PkgWatt against the power target. SysWatt=0.00 can mean the psys counter is absent or frozen; use PkgWatt for the effective measurement in that case.

For an orchestrated apply, monitor, stress, and report pass, use the combined-power-thermal-profiling skill. It enforces the same power-then- thermal sequence.

Restart Lifecycle#

The two tools manage state with different lifetimes:

State

Location

After OS restart

Required action

Package and psys RAPL PL1/PL2 caps

CPU MSRs and powercap state

Reset to firmware defaults.

Re-apply the intended power profile.

intel_lpmd tuning

intel_lpmd configuration files

Retained and read by intel_lpmd at boot.

No action for retention; restore the original configuration to remove it.

Thermal trip points and PPCC maximum

/etc/thermald/thermal-conf.xml

Retained and read when thermald starts.

Re-apply the thermal profile after the power profile to refresh PPCC with the new live PL1.

thermald sole-authority setting

/etc/systemd/system/thermald.service.d/override.conf

Retained.

No action unless restoring the packaged thermald behavior.

Monitoring and stress processes

Runtime processes and log file

Processes stop; log file remains.

Start a new monitor/test only when needed.

After every OS restart, re-apply the chosen profiles in this exact order:

sudo tools/power-tuning/set_power_profile.sh --profile BalancedHigh
sudo tools/power-tuning/set_thermal_profile.sh --profile warm

Do this even though the intel_lpmd and thermald files persist: the rebooted RAPL cap is volatile, and the thermal profile must be regenerated after the new cap is live. Re-run the second command after any future power-profile change or thermald restart that needs to use the new cap.

Restore Original Settings#

To stop using custom thermal control immediately, return thermald to kernel default thermal control:

sudo tools/power-tuning/set_thermal_profile.sh --disable

To restore the previous thermal files captured by the tool’s most recent apply, copy its backups back and restart thermald:

sudo cp -a /etc/thermald/thermal-conf.xml.bak /etc/thermald/thermal-conf.xml
sudo cp -a /etc/systemd/system/thermald.service.d/override.conf.bak /etc/systemd/system/thermald.service.d/override.conf
sudo systemctl daemon-reload
sudo systemctl restart thermald

Use the *.before-power-thermal copies made before the first apply when the goal is the original local configuration rather than only the previous profile. If thermald was originally disabled, stop and disable it again after restoring the files.

RAPL limits return to firmware defaults after a reboot. To restore the intel_lpmd daemon configuration immediately, restore the matching model-specific .orig file that the power tool created, then restart the daemon. For example:

sudo cp -a /usr/local/etc/intel_lpmd/intel_lpmd_config_F6_M204.xml.orig /usr/local/etc/intel_lpmd/intel_lpmd_config_F6_M204.xml
sudo systemctl restart intel_lpmd.service

The generic intel_lpmd_config.xml has no automatic backup when it did not exist before the first power-profile run. Restore a pre-change copy when one was captured, or use the platform’s known-good package configuration. Do not delete an existing configuration merely to approximate a reset.