Jetson Xavier NX Series and Jetson AGX Xavier Series

This topic describes power and performance management features of NVIDIA® Jetson Xavier™ NX series and NVIDIA® Jetson AGX Xavier™ series devices. It describes the power, thermal, and electrical management features visible to software, as well as some tools and related techniques.

Note

These devices’ power management features are very similar, and most of this document applies equally to both. For convenience, the text often refers to them by the shorter name Jetson Xavier. This term is used solely for convenience, and is not a product name.

Jetson Xavier and NVIDIA® Jetson™ Board Support Package (BSP) provide many features related to power management, thermal management, and electrical management. These features deliver the best user experience possible given the constraints of a particular platform. They help to create a user experience of:

  • Uniformly high performance

  • Excellent battery life

  • Perfect stability

  • Cool operation (the device is comfortable to touch)

Interacting Features

Power, thermal, and electrical management features place dynamic constraints on many operational settings (“knobs”), such as:

  • Clock gate settings

  • Clock frequencies

  • Power gate (or regulator enable) settings

  • Voltages

  • Processor power state (for example, which idle state is selected for the CPU)

  • Peripheral power state (for example, which idle state is selected for an I/O controller)

  • Chipset power state

  • Availability of CPU cores to the OS

Some of these knobs are constrained by more than one feature. For example, cpufreq implements load-based scaling, which adjusts the CPU frequency according to how busy the CPU is. CPU thermal management, however, can override the target frequency of cpufreq. Consequently, before you attempt to debug power, performance, thermal, or electrical problems, you must familiarize yourself with all of the power, thermal, and electrical management features in BSP.

Kernel Space Power Saving Features

This section describes BSP features that save power and extend battery life. Many of these features are implemented by the Linux kernel, with support from firmware and hardware, and without significant involvement from the user space.

Chipset Power States

The supported power states are listed in order of increasing flexibility or configurability:

  • Off: There is only one way for a system to be off.

  • Deep Sleep (SC7) offers a small amount of configurability. For example, before entering Deep Sleep, the software can select which hardware wake events can wake the chip from Deep Sleep.

  • Active state is extraordinarily flexible in terms of power and performance. It encompasses activity levels from low-power audio playback through peak performance. Power consumption in the Active state can range from tens of milliwatts to several watts.

Supported Power States

The supported power states are:

Power State

Characteristics

Off

Power rails: None of the power rails supplying the SoC and DRAM are powered.

State: No state is maintained in the SoC or DRAM.

Exiting: Into Active state via cold boot.

Deep Sleep (SC7)

Power rails: VDD_RTC, VDDIO_DDR, VDDIO_SYS, and DRAM power rails are powered on. VDD_CORE and VDD_CPU are powered off.

State: The SoC maintains a small amount of state information in the PMC block. DRAM maintains the state.

Exiting: Into Active state via a predefined set of wake events.

Active

Power rails: VDD_RTC, VDDIO_DDR, VDDIO_SYS, VDD_CORE, and DRAM rails are powered on. Other power rails, including VDD_CPU, may be powered on.

State: Software actively manages the power states of the devices that make up the SoC.

Exiting: Software can initiate a transition from Active to any other power state.

Power State Mapping to Linux

BSP maps chipset power states to Linux power states as follows.

Chipset power state

Linux power state

Comments

Off

Off

Deep Sleep (SC7)

Suspend to RAM

The software can choose whether to enter Deep Sleep before the OS enters Suspend.

Active

Running/Idle (display on or off)

Many SoC devices may be idle or disabled under driver control. For example, VDD_GPU may be powered off and the companion GPU may be power-gated.

Deep Sleep (SC7)

If the systemd init system is being used, you can initiate deep sleep from the user space with the following command:

$ sudo systemctl suspend

You can also use the following command:

$ sudo bash -c "echo mem > /sys/power/state"

The first method of entering deep sleep is preferred because it cooperates better with systemd, which maintains the Linux runlevel. If your system is not running systemd, use the second method.

The system can be awakened from deep sleep by common wake sources available on Jetson platforms:

Wake source

Usage

Power button

Press and release the power button on the Jetson device. If the power button is not available, connect and disconnect the power button pin and ground.

RTC alarm

Before entering low power state, program the RTC alarm with the following command:

$ sudo bash -c "echo +10 > /sys/class/rtc/rtc<x>/wakealarm"

where <x> is the RTC ID. The rtc<x> indicates the RTC used by system and can be found with the following command:

$ find /sys/class/rtc/* -maxdepth 0 -printf "%f:" -exec bash -c \
           "cat {}/hctosys" \; | grep :1 | cut -d: -f1 | head -n1

USB type-C cable hotplug

To flash the device, connect or disconnect a USB cable to the USB type-C port.

USB remote

Press any key on a USB keyboard connected to the device.

Wake on LAN

On another machine on the same LAN, enter:

$ sudo etherwake -i <interface> <MAC_address_of_target>

SD card detection

Insert or remove SD card.

Clock and Voltage Management

Because clock frequency is proportional to voltage, dynamic voltage scaling is closely related to frequency scaling. For example, higher clock frequencies require higher voltages and vice versa.

Most clock register manipulation on Jetson Xavier is handled by the Boot and Power Management (BPMP) firmware, which runs on the BPMP processor. A Linux kernel driver on the CPU exposes a somewhat simplified view of the physical clock tree to software on the main CPU via the Linux Common Clock Framework.

Each of the significant clock domains on the chip has its own dedicated clock source, known as a Noise Aware Frequency Lock Loop (NAFLL).

Regulator Framework

The Linux regulator framework provides an abstraction that allows regulator consumer drivers to dynamically adjust voltage or current regulators at runtime, without knowledge of the underlying hardware power tree.

The framework provides a mechanism that platform initialization code can use to declare a power tree topology and assign a driver that provides regulators for each node in the hardware power tree. Such a driver is called a regulator provider driver.

BSP configures the platform power tree appropriately for Jetson Xavier. Additionally, drivers within BSP act as regulator consumers, where appropriate.

When you port BSP to a new platform, you must ensure that:

  • The platform power tree is configured to match the underlying hardware.

  • All drivers for peripheral devices use the regulator consumer APIs correctly.

  • The device tree and board configuration file information for your new platform avoid conflicts between functions using the same I/O pads. BSP drivers registering as regulator consumers can cause I/O pads on the chip to be unavailable for other functions.

The SoC core power rails (VDD_CORE, VDD_CPU, VDD_GPU, and VDD_CV) are under the direct control of the BPMP firmware. They are configured via the BPMP device tree blob (which is distinct from the Linux device tree blob).

CPU Power Management

The CPU power management strategy uses dynamic frequency scaling (DFS) with dynamic voltage scaling, idle power states, and core management tuned for the Jetson Xavier architecture.

Frequency Management with cpufreq

BSP implements CPU dynamic frequency scaling with the Linux cpufreq subsystem. The cpufreq subsystem comprises:

  • Platform drivers to implement the clock adjustment mechanism

  • Governors to implement frequency scaling policies

  • A core framework to connect governors to platform drivers

The policy for frequency scaling depends on which cpufreq governor is selected at runtime.

For details, see the information at:

<top>/kernel/kernel-5.10/Documentation/admin-guide/pm/cpufreq.rst

For each Jetson platform, NVIDIA selects a cpufreq governor and tunes it to achieve a balance between power and performance.

When a governor requests a CPU frequency change, the cpufreq platform driver reconciles that request with constraints imposed by thermal or electrical limits, and updates the CPU clock speed.

Jetson Xavier uses an NAFLL to clock each CPU. The NAFLLs are configured for Adaptive Voltage and Frequency Scaling (AVFS). Hardware, with the assistance of the BPMP firmware, ensures that the CPU voltage is appropriate for the NAFLL to deliver requested CPU frequencies.

Idle Management with cpuidle

The Linux cpuidle infrastructure supports the implementation of SoC-specific idle states for each CPU core. cpuidle lacks direct support for idle states that are applicable to an entire CPU cluster or that extend beyond a CPU cluster.

For more information about the Linux cpuidle infrastructure, see:

<top>/kernel/kernel-5.10/Documentation/admin-guide/pm/cpuidle.rst

CPU Idle

NVIDIA provides an SoC-specific cpuidle driver that plugs into the cpuidle framework to enable CPU idle power management.

For each core there is an idle task that is scheduled when no other runnable tasks are left in that core’s run queue. This task places the core in a low-power state selected by the cpuidle governor. The core stays in that state until an interrupt wakes it up to process more work.

When the last active core in a CPU cluster goes into an idle or offline state, the idle task puts the entire CPU cluster in a low-power state.

Idle States

The table below summarizes the CPU core and cluster idle states available on Jetson Xavier, and the BSP software support for them.

Core states are denoted as Cx states, and cluster states are denoted as CCx states.

Type of state

State

Meaning

Software support?

Core state

C1

Clock gating

Yes

C6

Virtual retention (power gating and architecture state restored by MTS)

Yes *

C7

Power gating

No

Cluster state

CC1

Auto clock gating

Yes

CC3

fmax@Vmin or specified idle frequency

Yes

CC6

Cluster power gating (includes non-CPU logic)

Yes †

* C6 is disabled by default because the min-residency-us device tree property of C6 is set to 0xffffffff.

† Because C6 is disabled by default, there is no change to enter CC6. To enable C6/CC6, see To enable/disable a core/cluster power state at boot time.

To enable CPU idle

To enable CPU idle you must enable the appropriate kernel configuration option and the appropriate device tree node. Enabling either one alone is not effective.

  • To enable CPU idle in the configuration file, set the option:

    CONFIG_CPU_IDLE=y
    CONFIG_CPU_IDLE_TEGRA19X=y
    
  • To enable CPU idle in the device tree, enable the device tree node cpuidle:

    cpuidle {
        compatible = "nvidia,tegra19x-cpuidle";
        status = "okay";
    };
    

To disable cpuidle at boot time

  • Disable the device tree node cpuidle.

To display CPU idle status

  • Enter these commands to determine whether CPU idle is enabled by sysfs:

    $ cat /sys/devices/system/cpu/cpuidle/current_driver
    

    If CPU idle is enabled, the command displays:

    tegra19x_cpuidle_driver
    

To enable/disable a core/cluster power state at boot time

  • To enable a core/cluster power state, set the following properties of the appropriate core/cluster state node:

    • status to "okay"

    • min-residency-us to a reasonable value

    For example, to enable power state C6 with min-residency-us = 50000:

    c6 {
        compatible = "nvidia,tegra194-cpuidle-core";
        idle-state-name = "Virtual core powergate";
        wakeup-latency-us = <2000>;
        min-residency-us = <50000>;
        power = <0x3c>;
        pmstate = <0x6>;
        arm,psci-suspend-param = <0x6>;
        status = "okay";
        phandle = <0x26d>;
    };
    
  • To disable a core/cluster power state, use either of the following procedures. (Both procedures apply to the device tree tegra194-cpuidle.dtsi.)

    • Remove or disable the appropriate core/cluster state node.

    or

    • Modify the appropriate core/cluster state node by setting the min-residency-us property to a high value, e.g., 0xffffffff.

    For example, to disable power state C6:

    c6 {
        compatible = "nvidia,tegra194-cpuidle-core";
        idle-state-name = "Virtual core powergate";
        wakeup-latency-us = <2000>;
        min-residency-us = <0xffffffff>;
        power = <0x3c>;
        pmstate = <0x6>;
        arm,psci-suspend-param = <0x6>;
        status = "okay";
        phandle = <0x26d>;
    };
    

To get and set a CPU core’s power state

The pathnames of the nodes that represent core power states are:

/sys/devices/system/cpu/cpu<x>/cpuidle/state<y>

Where:

  • <x> is a core ID.

  • <y> is the index of the core power state: 0 for C1, or 1 for C6.

Note

A core power state’s status is 1 if the state is disabled, and 0 if it is enabledthe reverse of the usual Boolean sense of 0 and 1.

To get the status of core power state <y> on core <x>, read the appropriate node. To set the status, write an ASCII 0 to 1 to the node.

Following are several useful commands for getting and setting the core power state:

  • To display the name of the core power state with index <y>, enter the command:

    $ cat /sys/devices/system/cpu/cpu<x>/cpuidle/state<y>/name
    

    For example, this command displays the name of state0:

    $ cat /sys/devices/system/cpu/cpu<x>/cpuidle/state0/name
    
  • To get the status of core power state with index <y> on core <x>:

    $ cat /sys/devices/system/cpu/cpu<x>/cpuidle/state<y>/disable
    
  • To change the status of core power state with index <y> on CPU core <x>:

    $ echo <b> > /sys/devices/system/cpu/cpu<x>/cpuidle/state<y>/disable
    

To get cluster states

  • To get the status of the cluster states enabled for each cluster, read this node:

    /sys/kernel/debug/tegra_cpuidle/deepest_cc_state
    

The value returned is:

  • 1: Only CC1 is enabled

  • 6: CC1 and CC6 are enabled

To get the per-core state usage statistics

  • To get the number of times the kernel requested a specified core to enter a specified state, read this node:

    $ cat /sys/devices/system/cpu/cpu<x>/cpuidle/state<y>/usage
    
  • To get the number of times a specified core actually entered a specified state, enter the command:

    $ cat /sys/kernel/debug/tegra_mce/cstats
    

    For example, to get the number of times that core 2 has entered power state1, enter the command:

    $ cat /sys/devices/system/cpu/cpu2/cpuidle/state1/usage
    
  • To get the total time in microseconds that a specified core has spent in a specified state since boot, read the following device:

    $ cat /sys/devices/system/cpu/cpu<x>/cpuidle/state<y>/time
    

Memory Power Management

NVIDIA SoC chipsets include power saving features whose operation is largely invisible to software at runtime. Most of those features are statically enabled at boot, according to settings in the boot configuration table (BCT).

Additionally, BSP implements Dynamic Voltage and frequency scaling for the memory controller (EMC/MC) and DRAM to save power. The EMC BCT and DVFS table are specific to the board design. The EMC DVFS table is included in the platform BPMP device tree file.

EMC Frequency Scaling Policy

The following factors affect EMC frequency scaling policy at runtime:

  • The entries in the EMC DVFS table

  • The average memory bandwidth used (as measured by hardware)

  • Requests made by various device drivers (cpufreq, graphics drivers, USB, HDMI®, and display)

  • Any limits dynamically imposed by thermal throttling

Supported Modes and Power Efficiency

Jetson Xavier is designed with a high efficiency Power Management Integrated Circuit (PMIC), voltage regulators, and power tree to optimize power efficiency. It supports three optimized power budgets, such as 10 watts, 15 watts, and 30 watts. For each power budget, several configurations are possible with various CPU frequencies and number of cores online.

Capping the memory, CPU, and GPU frequencies, and number of online CPU, GPU TPC, DLA and PVA cores at a prequalified level confines the module to the target mode. Refer to the Thermal Design Guide, which you can find in the Jetson Download Center, for CPU / DLA / PVA heavy workloads. The configurations predefined by NVIDIA are as follows.

The MAXN mode is an unconstrained power mode that allows a maximum number of cores and clock frequency for CPU, GPU, DLA, PVA, and SOC engines like NVENC, NVDEC, and so on. However, this mode does not guarantee the best performance for all use cases because hardware throttling is engaged when the total module power exceeds the TDP budget. Therefore, it is not the maximum performance mode. This is an experimental mode to tweak clock settings and create custom power modes that balance performance and power consumption. Refer to Power Estimator for more information about estimating the power and generating the nvpmodel configuration file for the custom power mode.

NVP Model Clock Configuration for the Jetson Xavier NX Series

Property

Mode

15W

15W

15W

10W

10W

10W*

20W

20W

20W

Power budget

15W

15W

15W

10W

10W

10W

20W

20W

20W

Mode ID

0

1

2

3

4

5

6

7

8

Online CPUs

2

4

6

2

4

4

2

4

6

CPU maximal frequency (MHz)

1900

1400

1400

1500

1200

1900

1900

1400

1400

GPU TPC

3

3

3

3

3

3

3

3

3

GPU maximal frequency (MHz)

1100

1100

1100

800

800

510

1100

1100

1100

DLA cores

2

2

2

2

2

2

2

2

2

DLA maximal frequency (MHz)

1100

1100

1100

900

900

900

1100

1100

1100

PVA cores

1

1

1

1

1

1

1

1

1

PVA maximal frequency (MHz)

600

600

600

400

400

400

600

600

600

CVNAS maximal frequency (MHz)

576

576

576

460.8

460.8

460.8

576

576

576

Memory maximal frequency (MHz)

1600

1600

1600

1600

1600

1600

1866

1866

1866

SoC clocks maximal frequency (MHz) 10W and 15W modes

adsp: 300 nvcsi: 314 rce: 384
ape: 150 host1x: 204 sce: 345.6
axi_cbb: 204 isp: 576 se: 473.6
bpmp: 384 nvdec: 665.6 tsec: 371.2
bpmp_apb: 408 nvenc: 499.2 vi: 460.8
display: 600 nvjpg: 371.2 vic: 601.6
display_hub: 300 pex: 250

SoC clocks maximal frequency (MHz) 20W modes

adsp: 300 nvcsi: 314 rce: 384
ape: 150 host1x: 204 sce: 345.6
axi_cbb: 204 isp: 576 se: 704
bpmp: 384 nvdec: 793.6 tsec: 371.2
bpmp_apb: 408 nvenc: 729.6 vi: 460.8
display: 600 nvjpg: 460.8 vic: 601.6
display_hub: 300 pex: 250
* The default mode is 10W (mode ID 5).

NVP Model Clock Configuration for Jetson AGX Xavier and Jetson AGX Xavier 64GB

Property

Mode

MAXN

10W

15W

30W

30W

30W

30W

15W*

Power budget

n/a

10W

15W

30W

30W

30W

30W

15W

Mode ID

0

1

2

3

4

5

6

7

Online CPUs

8

2

4

8

6

4

2

4

CPU maximal frequency (MHz)

2265.6

1200

1200

1200

1450

1780

2100

2188

GPU TPC

4

2

4

4

4

4

4

4

GPU maximal frequency (MHz)

1377

520

670

900

900

900

900

670

DLA cores

2

2

2

2

2

2

2

2

DLA maximal frequency (MHz)

1395.2

550

750

1050

1050

1050

1050

115.2

PVA cores

2

0

1

1

1

1

1

1

PVA maximal frequency (MHz)

1088

0

550

760

760

760

760

115.2

CVNAS maximal frequency (MHz)

1356.8

601.6

716.8

1011.2

1011.2

1011.2

1011.2

115.2

Memory maximal frequency (MHz)

2133

1066

1333

1600

1600

1600

1600

1333

SoC clocks maximal frequency (MHz) All modes

adsp: 300 csi: 400 rce: 819.2
ape: 150 host1x: 408 sce: 729.6
axi_cbb: 408 isp: 1190.4 se: 1036.8
bpmp: 896 nvdec: 1190.4 tsec: 1036.8
bpmp_apb: 408 nvenc: 1075.2 vi: 998.4
display: 800 nvjpg: 716.8 vic: 1036.8
display_hub: 400 pex: 500
* The default mode is 15W (mode ID 7).
Default mode is intended to improve desktop application performance.
PVA and DLA are not in use and run at minimal frequency.

NVP Model Clock Configuration for Jetson AGX Xavier Industrial

Property

Mode

MAXN

20W

40W*

40W

40W

40W

Power budget

n/a

20W

40W

40W

40W

40W

Mode ID

0

1

2

3

4

5

Online CPUs

8

4

8

6

4

2

CPU maximal frequency (MHz)

2035.2

1200

1200

1450

1780

2035.2

GPU TPC

4

4

4

4

4

4

GPU maximal frequency (MHz)

1211.3

670

900

900

900

900

DLA cores

2

2

2

2

2

2

DLA maximal frequency (MHz)

1228.8

750

1050

1050

1050

1050

PVA cores

2

1

1

1

1

1

PVA maximal frequency (MHz)

947.2

550

760

760

760

760

CVNAS maximal frequency (MHz)

1203.2

716.8

1011.2

1011.2

1011.2

1011.2

Memory maximal frequency (MHz)

2133

1600

1600

1600

1600

1600

SoC clocks maximal frequency (MHz) All modes

adsp: 300 csi: 400 rce: 678.4
ape: 150 host1x: 408 sce: 588.8
axi_cbb: 408 isp: 1011.2 se: 857.6
bpmp: 755.2 nvdec: 960 tsec: 806.4
bpmp_apb: 408 nvenc: 870.4 vi: 819.2
display: 768 nvjpg: 563.2 vic: 819.2
display_hub: 358.4 pex: 500
* The default mode is 40W (mode ID 2).

Power Mode Controls

You can display and change the power mode with the nvpmodel command.

  • To change the power mode, enter the command:

    $ sudo /usr/sbin/nvpmodel -m <x>
    

    Where <x> is the power mode ID (i.e. 0, 1, 2, 3, 4, 5, or 6).

    Alternatively, use the nvpmodel GUI front end. For more information, see nvpmodel GUI, later in this topic.

    After you set a power mode, the module stays in that mode until you change it. The mode persists across power cycles and SC7.

Note

GPU tpc_pg_mask can be set once before the GPU golden context is created. If the nvpmodel power mode change requires to set the different tpc_pg_mask value then the system reboot is required.

  • Example:

    ubuntu@jetson:~$ sudo nvpmodel -m 1
    [351953.975356] nvgpu: 17000000.gv11b tpc_pg_mask_store:1053 [ERR]  golden image size already initialized
    NVPM ERROR: Error writing 5 to /sys/devices/gpu.0/tpc_pg_mask: 19
    NVPM WARN: Reboot required for changing to this power mode: 1
    NVPM WARN: DO YOU WANT TO REBOOT NOW? enter YES/yes to confirm:
    

Type YES or yes to initiate reboot or press any other key to cancel. The settings will be in effect after the reboot.

Ignore the NVPM ERROR: Error writing 5 to /sys/devices/gpu.0/tpc_pg_mask: 19 statement as this is just for the information while change in GPU tpc_pg_mask.

  • To display the current power mode, enter the command:

    $ sudo /usr/sbin/nvpmodel -q
    

    Alternatively, see the mode displayed to the right of the NVIDIA icon in the nvpmodel window’s menu bar. For more information, see nvpmodel GUI, later in this topic.

  • To add a custom power mode definition, edit this file:

    /etc/nvpmodel.conf
    

    This is an example entry for mode 2:

    < POWER_MODEL ID=2 NAME=MODE_15W >
    CPU_ONLINE CORE_0 1
    CPU_ONLINE CORE_1 1
    CPU_ONLINE CORE_2 1
    CPU_ONLINE CORE_3 1
    CPU_ONLINE CORE_4 0
    CPU_ONLINE CORE_5 0
    CPU_ONLINE CORE_6 0
    CPU_ONLINE CORE_7 0
    CPU_DENVER_0 MIN_FREQ 1200000
    CPU_DENVER_0 MAX_FREQ 1200000
    CPU_DENVER_1 MIN_FREQ 1200000
    CPU_DENVER_1 MAX_FREQ 1200000
    GPU MIN_FREQ 0
    GPU MAX_FREQ 670000000
    EMC MAX_FREQ 1331200000
    DLA_CORE MAX_FREQ 750000000
    DLA_FALCON MAX_FREQ 450000000
    PVA_VPS MAX_FREQ 550000000
    PVA_CORE MAX_FREQ 385000000
    

    The unit of measure for CPU frequency is kilohertz. The unit for GPU and EMMC frequency is hertz. You must assign each custom mode a unique number in the ID field. Test your use case to determine:

    • How many active cores to use

    • The frequency for each CPU cluster, and the GPU and EMC frequencies

    The frequencies you select are subject to the MaxN limit defined in mode 0.

  • To learn about other options, enter the command:

    $ /usr/sbin/nvpmodel -h
    

Fan Profile Control

Jetson Xavier supports two profiles of fan operation named “quiet” and “cool.”

Userspace fan speed control daemon nvfancontrol manages fan speed based on the trip point temperatures configured for the selected profile.

Fan Profile Configuration

Every fan speed step is associated with the trip point temperature and corresponding hysteresis. The following table shows the configurations predefined by NVIDIA.

Fan Profile Configuration for the Jetson Xavier NX Series

Fan profile "quiet"

Trip temperature*

0

46

60

68

76

Hysteresis*

0

8

8

7

7

Fan PWM value

0

130

160

200

255

Fan profile "cool"

Trip temperature*

0

35

45

53

61

Hysteresis*

0

8

8

7

7

Fan PWM value

0

140

170

200

255

* Trip temperature and hysteresis in degrees Celsius.

Fan Profile Configuration for the Jetson AGX Xavier Series

Fan profile "quiet"

Trip temperature*

0

50

63

72

81

Hysteresis*

0

18

8

8

8

Fan PWM value

0

77

120

160

255

Fan profile "cool"

Trip temperature*

0

35

53

62

73

Hysteresis*

0

9

8

8

9

Fan PWM value

0

77

120

160

255

* Trip temperature and hysteresis in degrees Celsius.

The framework implements hysteresis to prevent frequent changes in fan speed. For Jetson Xavier, as an example, when fan profile is set to “quiet” with the default settings shown above, the framework performs these actions:

  • Turns on the fan when the temperature rises to 50 °C

  • Turns off the fan when the temperature falls to 32 °C

  • Turns on the fan again when the temperature rises to 50 °C, and so on

nvfancontrol

nvfancontrol is a userspace fan speed control daemon. This manages the fan speed based on the temperature-to-fan-speed mapping table in the nvfancontrol configuration file. There are some basic elements in the nvfancontrol service, including Tmargin, kickstart PWM, fan profile, fan control, and fan governor. All of these can be programmed via the configuration file based on the user’s preferences. This chapter will explain each of them in the following sections.

nvfancontrol.conf

  • Location:

    /etc/nvfancontrol.conf
    
  • Please find below the example nvfancontrol.conf file for the Jetson Xavier:

    POLLING_INTERVAL 2
    <FAN 1>
        TMARGIN DISABLED
        FAN_GOVERNOR pid {
                STEP_SIZE 10
        }
        FAN_PROFILE quiet {
                #TEMP   HYST    PWM     RPM
                0       0       0       0
                50      18      77      1000
                63      8       120     2000
                72      8       160     3000
                81      8       255     4000
                140     0       255     5000
                150     0       255     6000
                160     0       255     7000
                170     0       255     10000
                180     0       255     11000
        }
        FAN_PROFILE cool {
                #TEMP   HYST    PWM     RPM
                0       0       0       0
                35      9       77      1000
                53      8       120     2000
                62      8       160     3000
                73      9       255     4000
                140     0       255     5000
                150     0       255     6000
                160     0       255     7000
                170     0       255     10000
                180     0       255     11000
        }
        THERMAL_GROUP 0 {
                GROUP_MAX_TEMP 108
                #Thermal-Zone Coeffs Max-Temp
                CPU-therm 30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
                GPU-therm 30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
                AUX-therm 40,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
        }
        FAN_DEFAULT_CONTROL open_loop
        FAN_DEFAULT_PROFILE quiet
        FAN_DEFAULT_GOVERNOR pid
    

Default Fan Profile

For Jetson Xavier the fan profile is set to “quiet” by default. It is defined as FAN_DEFAULT_PROFILE in the configuration file /etc/nvfancontrol.conf.

To change the default fan profile

To change the fan profile, complete the following steps:

  • Stop the nvfancontrol systemd service:

    sudo systemctl stop nvfancontrol
    
  • Set the default fan profile by putting the following property in /etc/nvfancontrol.conf:

    FAN_DEFAULT_PROFILE <fan_profile>
    

Where <fan_profile> is quiet or cool.

  • Remove the status file:

    sudo rm /var/lib/nvfancontrol/status
    
  • Start the nvfancontrol systemd service:

    sudo systemctl start nvfancontrol
    

To identify the current fan profile

  • Enter the command:

    $ sudo nvfancontrol -q
    
  • Example:

    $ sudo nvfancontrol -q
    FAN1:FAN_PROFILE:quiet
    ...
    ...
    

Once you set a fan profile, the module stays in that profile until you change it. The profile persists across power cycles and SC7.

Fan Profile Table

The fan profile table contains the mapping between temperature and the fan speed. It also contains the hysteresis value for each step and the fan RPM value.

  • Syntax:

    FAN_PROFILE <fan_profile_name> {
            <temp>  <hyst>  <pwm>   <rpm>
    }
    
    Where:
    <fan_profile_name>: Fan Profile Name
    <temp>: Temperation step in degree celcius
    <hyst>: Hysteresis step
    <pwm>:  Fan PWM value
    <rpm>:  Fan RPM value
    
  • Example:

    FAN_PROFILE quiet {
            #TEMP   HYST    PWM     RPM
            0       0       0       0
            50      18      77      1000
            63      8       120     2000
            72      8       160     3000
            81      8       255     4000
            140     0       255     5000
            150     0       255     6000
            160     0       255     7000
            170     0       255     10000
            180     0       255     11000
    }
    

TMARGIN

TMARGIN temperature is the difference between the maximum allowable temperature and the current thermal zone temperature. For example, if the maximum allowable temperature of CPU-therm is 105 degree Celsius, and the current temperature of CPU-therm is 45 degree Celsius, the current TMARGIN temperature of CPU-therm is 60 degree Celsius (105 - 45).

Kickstart PWM

The minimal required PWM value to start the fan from complete stop state is called kickstart PWM. The fan might not start spinning if PWM value is lower than kickstart PWM.

Thermal Group

THERMAL_GROUP contains the list of thermal zones considered for calculating the trip temperature and the group max temperature for calculating the TMARGIN temperature.

This section contains:

  • Thermal Group Max temperature:

    GROUP_MAX_TEMP <temp_in_degree_celcius>
    

This parameter is used only in case when TMARGIN is ENABLED. The Tmargin temperature is calculated as shown in TMARGIN section.

  • Thermal zone name, coefficients and the thermal zone max temperature:

    <thermal_zone_name> <coeff_0>,<coeff_2>....,<coeff_19> <thermal_zone_max_temp>
    
    Where:
    <thermal_zone_name>: Thermal zone name
    <coeff_0..coeff_19>: Coefficients used for calculating weighted average. Please note that
                         only <coeff_0> is taken into consideration as of now.
    <thermal_zone_max_temp>: Thermal zone max temperature. This is used only when Tmargin
                             is ENABLED. If GROUP_MAX_TEMP specified then this temperature
                             will be ignored.
    
  • Here is an example for calculating weighted average temperature with Tmargin disabled. Consider that the current CPU-therm is 40 degree Celsius, GPU-therm is 40 degree Celsius, and AUX-therm is 38 degree Celsius, then weighted average temperature = 40 * 0.3 + 40 * 0.3 + 38 * 0.4 = 39 degree Celsius with below thermal group:

    THERMAL_GROUP 0 {
            GROUP_MAX_TEMP 105
            #Thermal-Zone Coeffs Max-Temp
            CPU-therm 30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
            GPU-therm 30,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
            AUX-therm 40,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 0
    }
    

Fan Control

Here are the fan control types in the nvfancontrol service:

  • open-loop:

    The open-loop fan control adjusts the fan speed by setting the desired PWM value based on the current trip temperature step. The RPM values in the profile are ignored.

  • closed-loop:

    The closed-loop fan control makes the fan spin close to the desired RPM value based on the current trip temperature step. The PWM values in the profile are ignored.

    To have the fan spin at the exact same speed as the target RPM, there will be a performance drop and the risk of shorter fan life due to the constant adjustment of the speed. So, there is a programmable tolerance value between the target RPM and the current RPM value (an rpm difference within 100 is acceptable in the below example):

    FAN_CONTROL close_loop {
            RPM_TOLERANCE 100
    }
    

Fan Governor

The fan governor decides the fan speed control logic based on the fan profile. There are two kinds of fan governors in the nvfancontrol service: pid and cont.

  • The following profile with Tmargin disabled and open-loop control can be used to explain how different fan governors handle it:

    TMARGIN DISABLED
    FAN_PROFILE quiet {
            #TEMP   HYST    PWM     RPM
            0       0       0       0
            50      18      77      1000
            63      8       120     2000
            72      8       160     3000
            81      8       255     4000
            140     0       255     5000
            150     0       255     6000
            160     0       255     7000
            170     0       255     10000
            180     0       255     11000
    }
    
  • pid:

    • The pid governor will change the fan speed only when the weighted average temperature crosses the trip temperature step. The curve between the weighted average temperature and fan speed resembles a stair.

    • For example, when a weighted average temperature increasing, and the current weighted average temperature is 50 degree Celsius, the PWM will be set to 77. Later even when the weighted average temperature goes to 62 degree Celsius, the PWM will still be set to 77. When the weighted average temperature goes to 63 degree Celsius, the PWM will be set to 120 until the next trip temperature step is crossed.

  • cont:

    • The cont governor will linearly interpolate the fan speed based on the upper and lower fan speed between the trip temperature steps. Compared to the pid governor, the curve between temperature and fan speed is more continuous.

    • For example, when the current weighted average temperature is 55 degree Celsius, the PWM will be set to 93 (77 + (55 - 50) * (120 - 77) / (63 - 50)).

Hysteresis in nvfancontrol

In nvfancontrol, hysteresis is used to define the fan speed change temperature threshold when using pid governor.

  • Consider below profile with Tmargin enabled:

    TMARGIN ENABLED
    FAN_PROFILE cool {
            #TEMP   HYST    PWM     RPM
            0       0       255     2900
            18      9       255     2900
            30      11      202     2300
            45      11      149     1700
            60      14      88      1000
            105     0       0       0
    }
    

The fan will be turned on when the Tmargin temperature reaches 60 degree Celsius. Later when the Tmargin temperature keeps increasing, after the Tmargin temperature goes over 74 degree Celsius (60 + 14 = 74) the fan will be turned off.

Polling Interval

nvfancontrol daemon polls the thermal zone temperatures at the time interval specified by POLLING_INTERVAL and sets the fan speed value specified as per TEMPERATURE - FAN SPEED mapping table.

  • Syntax:

    POLLING_INTERVAL <time_in_seconds>
    

TMARGIN Configuration

TMARGIN configuration needs to be specified in order for nvfancontrol daemon to depict the Fan Profile Table correctly.

  • TMARGIN ENABLED Case:

  • Formula to calculate the Tmargin temperature:

    Tmargin_sensor_temp = GROUP_MAX_TEMP -OR- <thermal_zone_max_temp> - <current_thermal_zone_temp>
    
  • Formula to calculate Tmargin weighted average of the thermal group sensors:

    Tmargin_thermgroup_weighted_average = Tmargin_sensor0_temp * sensor0_weight_ratio + Tmargin_sensor1_temp * sensor1_weight_Ratio + ...
    
    Where:
    Tmargin_sensor<x>_temp - Tmargin sensor temperature calculated using above formula.
    sensor<x>_wight_ratio - Currently only <coeff_0> value is considered for weight ratio as mentioned in section "Thermal Group"
    x - sensor number
    
  • TMARGIN ENABLED Table:

        TMARGIN ENABLED
        FAN_PROFILE quiet {
                #TEMP   HYST    PWM     RPM
                0       0       255     0
                50      18      160     1000
                63      8       120     2000
                72      8       77      3000
                100     8       0       4000
        }
    
    Example:
    
    Temperature steps defined in the above table are the Tmargin temperatures
    calculated using the formula mentioned at the start of this section.
    
    Consider that the GROUP_MAX_TEMP is set as 108, the current fan governor is
    pid, and the current fan control is open-loop. So, as specified in the above
    temperature-to-fan-speed mapping table, the Tmargin trip temperature step
    63 degree Celsius corresponds to (108 - 63) = 45 degree Celsius, which is
    the weighted average of the thermal zone temperature.
    
    When the weighted average of the thermal zone temperature reaches 46 degree
    Celsius, then nvfancontrol sets the fan PWM to 160. The fan PWM will stay
    at 160 till the weighted average of the thermal zone temperature reaches
    58 degree Celsius (108 - 50 = 58) which is the Tmargin trip point step 50 degree
    Celsius as present in the above table.
    
  • TMARGIN DISABLED Case:

  • Formula to calculate weighted average of the thermal group sensors:

    thermgroup_weighted_average = sensor0_temp * sensor0_weight_ratio + sensor1_temp * sensor1_weight_Ratio + ...
    
    Where:
    sensor<x>_temp - Current thermal zone temperature.
    sensor<x>_wight_ratio - Currently only <coeff_0> value is considered for weight ratio as mentioned in section "Thermal Group"
    x - sensor number
    
  • TMARGIN DISABLED Table:

        TMARGIN DISABLED
        FAN_PROFILE quiet {
                #TEMP   HYST    PWM     RPM
                0       0       0       0
                50      18      77      1000
                63      8       120     2000
                72      8       160     3000
                81      8       255     4000
        }
    
    Example:
    
    Temperature steps defined in the above table are the weighted average
    of the actual thermal zone temperature.
    
    Consider that the current fan governor is pid, and the current fan control
    is open-loop. As specified in the above temperature-to-fan-speed mapping
    table, when the actual weighted average of the thermal zone temperature
    reaches 50 degree Celsius, then nvfancontrol sets the fan PWM to 77.
    The fan PWM will stay at 77 until the weighted average of the thermal
    zone temperature reaches 63 degree Celsius and so on.
    

Thermal Management

Thermal management is essential for system stability and quality of user experience. Jetson Xavier thermal management provides the following capabilities:

  • Sensing for on-board and on-chip thermal sensor temperature reporting

  • Cooldown for removing heat via the fan and for controlling heat via software clock throttling

  • Slowdown for hardware clock throttling

  • Shutdown for orderly software shutdown and hardware thermal shutdown

Thermal management in Jetson Xavier is performed by:

  • The Linux kernel, which monitors on-board thermal sensors, performs cooldown, and supports software and hardware thermal shutdown

  • The Board and Power Management Processor (BPMP), which monitors on-chip thermal sensors, and performs slowdown and hardware thermal shutdown

The following table identifies each thermal management action and the associated module for the SoC.

Thermal Action

Linux Device Driver

Associated Module

Sensing

soctherm.c

BPMP firmware

aotag.c

BPMP firmware

nct1008.c

Kernel software

Cooldown for software throttling

cpufreq_cooling.c devfreq_cooling.c

Kernel software

Cooldown for fan

pwm_fan.c

Kernel software

Slowdown for hardware throttling

soctherm.c

BPMP firmware

Software shutdown

thermal_core.c

Kernel software

Hardware shutdown

soctherm.c and aotag.c

BPMP firmware

nct1008.c

Kernel software

Linux Thermal Framework

The Linux thermal framework provides generic user space and kernel space interfaces for working with devices that measure or control temperature. The central component of the framework is the thermal zone.

For more information about the Linux thermal framework, see:

<top>/kernel/kernel-5.10/Documentation/driver-api/thermal/sysfs-api.rst

Thermal Zone

A thermal zone is a virtual object that represents an area on the die whose temperature is monitored and controlled. A thermal zone acts as an object with the following components:

  • Temperature sensor

  • Cooling device

  • Trip points

  • Governor

BSP includes drivers that provide interfaces to these components.

This topic introduces these components and demonstrates how they form a thermal zone on a Jetson device.

Configuring a Thermal Zone Using the Device Tree

A thermal zone provides knobs to tune the thermal response of the zone. BSP provides several thermal zones tuned to provide optimum thermal performance. You can modify the provided thermal zones by editing the entries in the kernel device tree. Users can define sensors to use temperature limits and cooling actions on those limits. Device overheating can be resolved in most cases by tuning the thermal zone.

The following code snippet provides an example of a thermal zone for Jetson Xavier. This thermal zone monitors the temperature of the THERMAL_ZONE_GPU sensor. Clock throttling is performed using the devfreq cooling device when the passive trip point, gpu-sw-throttle, is crossed at 92.5 °C:

GPU-therm {
                polling-delay = <0x0>;
                polling-delay-passive = <0x1f4>;
                thermal-sensors = <0x28d 0x3>;
                status = "okay";

                thermal-zone-params {
                        governor-name = "step_wise";
                };

                trips {

                        trip_critical {
                                temperature = <0x17ed0>;
                                type = "critical";
                                hysteresis = <0x0>;
                                writable;
                        };

                        gpu-sw-throttle {
                                temperature = <0x16954>;
                                type = "passive";
                                hysteresis = <0x0>;
                                writable;
                                phandle = <0x298>;
                        };
                };

                cooling-maps {

                        map0 {
                                trip = <0x298>;
                                cooling-device = <0x299 0xffffffff 0xffffffff>;
                        };

                        user-alert-map0 {
                                trip = <0x298>;
                                cooling-device = <0x29a 0x1 0x1>;
                        };
                };
        };

For more information about thermal knobs, see:

<top>/kernel/kernel-5.10/Documentation/devicetree/bindings/thermal/nvidia,tegra186-bpmp-thermal.txt

Temperature Sensors

A temperature sensor in a thermal zone is responsible for reporting the temperature in millidegrees Celsius. Jetson Xavier has several types of temperature sensors on the chip and board.

For more information see Thermal Sensing in Linux.

Trip Points and Cooling Devices

Thermal management uses trip points to communicate with thermal zones. A trip point describes the temperature at which cooling is recommended.

Trip points are classified by the type of cooling device is triggered:

  • Passive trip points trigger passive cooling devices, which reduce the Jetson device’s performance, and so reduce the amount of heat generated. Hardware or software clock throttling (reducing the frequency of a clock) is an example of a passive cooling device.

  • Active trip points trigger active cooling devices, which activates a device to remove the dissipated heat. A fan is an example of an active cooling device.

  • Critical trip points trigger a thermal shutdown.

A cooling map specifies how a cooling device is associated with certain trip points.

For more information, see Thermal Cooling.

Governors

A governor implements a feedback control loop that keeps a Jetson device within a safe operating temperature range. While the Linux thermal framework provides many different governors, BSP provides a simple Proportional Integral Derivative (PID) controller for all passive throttling needs.

BSP-Specific Thermal Zones

BSP defines platform-specific thermal zones. The zones are tuned to provide the best performance within the thermal constraints of the Jetson device. Each thermal zone uses a temperature sensor that is controlled by either the Linux kernel or the BPMP firmware, as described in the following table.

Thermal Zone

Thermal Sensor

Associated Module

CPU-therm

THERMAL_ZONE_PLLX

BPMP firmware

GPU-therm

THERMAL_ZONE_GPU

BPMP firmware

AUX-therm

THERMAL_ZONE_AUX

BPMP firmware

AO-therm

THERMAL_ZONE_AO

BPMP firmware

Tdiode_tegra

tmp451

Linux kernel

PMIC-Die

Power management integrated circuit (PMIC)

Linux kernel

Tboard_tegra

tmp451

Linux kernel

For more information, see Thermal Management in BPMP.

Gains achieved by tuning are limited by the Thermal Design Power (TDP) of the system. Tuning cannot remedy a faulty TDP. Removing all of the thermal zones does not guarantee maximum performance, and can cause resets and/or irreversible damage to the device.

Thermal Management in Linux

The Linux kernel provided by BSP includes several drivers for on-board and on-chip temperature sensing.

Thermal Sensors

Jetson Xavier series has several types of sensors to support hardware and software cooling strategies.

On-board Sensors

BSP includes a driver for on-board sensor devices such as:
  • NCT1008

  • NCT72

  • TMP451

Note

Jetson Xavier NX does not have any on-board sensor.

These devices can sense their own temperature as well as the temperature of a remote diode. Jetson platforms have these sensors set up as follows:

Thermal Zone

Thermal Sensor

Sensed Location

Tdiode_tegra

Remote sensor

Temperature on die near GPU

Tboard_tegra

Local sensor

Temperature of the board

BSP configures these sensors to operate in an extended mode to increase the temperature range to −64 °C to 191 °C.

Operation in SC7

The voltage rail that powers the on-board sensor is gated when the SoC enters the SC7 state on most Jetson Platforms. Consequently, the sensor is stopped when the SoC enters SC7 and restarted when it exits that state.

Thermal Capabilities

The On-board sensors generate thermal events for:

  • Thermal zone trip points

  • Hardware thermal shutdown

Correction Offset

The on-board sensors allow software to program a static offset temperature for the remote sensor. This accounts for any inaccuracy that may be present in the sensor hardware. BSP reads the offset from the device tree and programs it into the offset register on boot. The offset is calculated and validated via oil bath experiments.

On-chip Sensors

The on-chip SOC_THERM and AOTAG thermal sensors are controlled by BPMP firmware and tegra-bpmp-thermal Linux kernel driver.

The BPMP firmware exposes each on-chip thermal sensor using the Application Binary Interface (ABI), and has an ABI name shown in the table in BSP-Specific Thermal Zones. The on-chip sensors, with the THERMAL_ZONE prefix, work as described in the following paragraphs.

The BPMP firmware has one programmable temperature threshold (one trip) for each on-chip sensor, allocated for a Linux thermal zone trip point. The tegra_bpmp_thermal driver walks through the list of thermal trip points in a Linux thermal zone based on the current temperature. It then comes up with a trip to program the sensor temperature threshold in BPMP firmware. The driver then uses the following thermal message requests (MRQs) to communicate with the BPMP thermal framework.

  • CMD_THERMAL_QUERY_ABI

  • CMD_THERMAL_GET_TEMP

  • CMD_THERMAL_SET_TRIP

  • CMD_THERMAL_GET_NUM_ZONES

The driver receives a CMD_THERMAL_HOST_TRIP_REACHED MRQ message when a particular sensor crosses a trip. The message is then relayed back to the Linux thermal framework.

Jetson Xavier termal framework architecture

For more information on thermal management features provided as part of BSP, see Thermal Management in BPMP.

Thermal Cooling

BSP provides thermal management using fan control and throttling of various clocks in the system.

Fan Management

BSP provides active cooling by fan management through the pwm-fan driver, controlled by nvfancontrol, which provides:

  • Fan speed control by programming the PWM controller

  • Ramp-up and ramp-down control to change the speed of the fan smoothly

  • Fan control during various power states

SoC thermal management uses the fan as the first line of defense to delay clock throttling until a much higher temperature is reached.

Software Clock Throttling

BSP provides thermal cooling by throttling various clocks in the system. When a thermal sensor’s temperature rises above a throttling trip point, clock throttling employs the DVFS capabilities of the clocks to reduce their operating frequencies, and thereby the voltages of the rails that power the clocks. This reduction in frequency and voltage reduces power consumption which helps to control the temperature.

Because BSP provides cooling by reducing the clock frequency, it directly impacts performance and the user experience. If a device feels warm and seems sluggish, it may be due to thermal throttling on the clocks. This can be remedied by tuning the trip points and cooling devices of thermal zones.

BSP Provides following cooling devices for software clock throttling:

  • cpufreq_cooling

  • devfreq_cooling

Each of these cooling devices provides several cooling states, each of which translates to a maximum allowable operating frequency for the CPU, GPU, and EMC clocks. These frequencies are optimized to provide the best possible performance at a given temperature. The frequency tables for these clocks are part of the device tree binary.

The governor uses the current temperature of a thermal zone as an input to the feedback control loop. Similarly, it uses the output of the control loop to set a new cooling state for the thermal zone’s cooling device. As the device heats up the governor picks progressively higher cooling states, which result in higher frequency caps for all of the clocks, and potentially greater cooling. BSP performs this thermal throttling of the clocks to maintain the junction temperature of the die within the recommended safe limits. For software throttling trip temperatures, see the table in Thermal Specifications.

Software Thermal Shutdown

A critical trip point triggers a software thermal shutdown. It allows the operating system to save its state and perform an orderly shutdown before a hardware thermal reset occurs.

A software thermal shutdown is considered a rare event. It occurs after all other cooling strategies have failed.

BSP defines one critical trip point per thermal zone. You can set the lower limit for the orderly shutdown. For software thermal shutdown trip temperatures, see the table in Thermal Specifications.

Hardware Thermal Shutdown

The on-chip and on-board sensors can trigger hardware shutdown when all other cooling strategies have failed, and software shutdown has failed to occur when it should. For hardware shutdown limits, see the table in Thermal Specifications.

Thermal Management in BPMP

BSP thermal management features are part of the firmware running on BPMP for Jetson platforms running any host operating system (host OS) on the CPU.

Thermal Sensing

The BPMP firmware hosts the soctherm and aotag drivers for the on-chip thermal sensors as follows:

Thermal Sensor

ABI Name

Sensed Location

AOTAG

AOTAG

THERMAL_ZONE_AO

Co-locate with TDIODE in pad-ring

SOC_THERM

PLLX

THERMAL_ZONE_PLLX

Center of CPU cluster

AUX (x3)

THERMAL_ZONE_AUX

Near CV cluster, SoC cluster

CPU

THERMAL_ZONE_CPU

Center of CPU cluster

GPU

THERMAL_ZONE_GPU

Center of GPU

SOC_THERM

SOC_THERM is the collection of on-chip ring oscillators whose frequency changes are based on temperature. To convert a measured frequency to a temperature, the oscillating frequency of the sensor, at a fixed temperature, must be known in advance and stored in the on-chip fuses.

The BPMP firmware soctherm driver uses these fuses during boot and calibrates the sensor. Once the calibration is complete, the temperature sensor reports the temperature, in degrees Celsius, with a 0.5 °C precision margin.

Sensors and Sensor Groups

The temperature sensors on the chip are logically classified in sensor groups, based on their proximity to certain hardware blocks. The sensor groups are represented as a single sensor to the host OS and the BPMP firmware.

For example, Jetson Xavier has two temperature sensors in the SOC cluster and one near the CV cluster. These are grouped as AUX sensors that are represented as THERMAL_ZONE_AUX to the operating system running on the CPUs. The BPMP firmware reports the temperature of a given group by taking the maximum of all the sensors in the group.

Thermal Event Detection

Thermal sensors can report the temperature when the current temperature crosses a software-defined trip point. The sensors are capable of monitoring several of these software trip points to perform the following thermal actions:

  • Report when the thermal trip point has been crossed

  • Trigger a hardware thermal shutdown

  • Trigger hardware throttling

Voltage Rail Dependencies

To provide accurate temperature sensing, the sensors require a minimum voltage. Additionally, the sensors cannot operate when the rail is power-gated.

When the system is in a low-power state, the firmware provides the following modes of operation:

  • No temperature measurements during SC7: Because the rail powering the sensor is power-gated in the SC7 state, the oscillator is not running. Therefore, the frequency-to-temperature conversion may result in inaccurate values. To avoid spurious temperature reports from the sensors, stop the sensors before entering the SC7 state.

    The firmware provides the AOTAG sensor for measuring temperature in the SC7 state. When the SC7 state is exited, the sensors are restarted.

  • Fallback to PLLX sensor: To ensure accurate temperature readings during minimum voltage, use the PLLX sensor’s oscillator. On platforms where the minimum voltage is not guaranteed, the firmware falls back on the PLLX sensor’s oscillator with a programmable offset. The result is that all the sensors invalidate their oscillators and use the PLLX sensor’s oscillator with the added offset. This fallback on the PLLX sensor’s oscillator allows for continuous temperature measurement, even at lower voltage levels.

As a side effect of PLLX fallback, the programmable offset compensates for the fact that the PLLX sensor’s oscillator is farther away than the oscillator that it is replacing. The host OS continues to use all of the thermal zones without side effects. The offset ensures that the CPU sensor reports more accurate temperatures than the PLLX sensor. The host OS must therefore continue to use the right sensors for measuring the CPU temperatures.

AOTAG

The Always-On Thermal Alert Generator (AOTAG) is a ring oscillator based temperature sensor. It is in the always-on power domain and can monitor temperatures even when the device is in the SC7 state. Apart from this distinction, the AOTAG sensor operates the same as any of the SOC_THERM sensors.

Thermal Event Detection

Just like the SOC_THERM sensor, the AOTAG sensor can generate interrupts. Additionally, it can monitor two software-controlled interrupt levels that BSP uses as:

  • Thermal zone trip points

  • Hardware thermal shutdown

BPMP Thermal Framework

The BPMP firmware hosts a thermal framework to:

  • Register thermal sensors as thermal zones as identified in Thermal Sensing

  • Allow BPMP modules to register trip points on the thermal zones

  • Allow the host OS to register trips using thermal MRQ messages

  • Provide trip management and reporting

The thermal framework maintains a list of trips per sensor that includes the current trip from the host OS and various BPMP modules. As temperatures change, the framework examines the list of current trips and notifies the owners of the trips of the changes. The notification is sent using a callback for the BPMP owned trips and the thermal MRQ command CMD_THERMAL_HOST_TRIP_REACHED for trips that are owned by the host OS.

The primary thermal MRQ requests handled by the framework are:

  • CMD_THERMAL_QUERY_ABI

  • CMD_THERMAL_GET_TEMP

  • CMD_THERMAL_SET_TRIP

  • CMD_THERMAL_GET_NUM_ZONES

Since there can be several trips on a given sensor, the thermal framework must ensure that a notification is generated whenever a given trip is crossed. For example, if THERMAL_ZONE_CPU has trips at 55°, 60°, 65°, and 70 °C, the thermal framework sends a single notification when the temperature crosses 55°, 60°, 65°, and 70 °C.

Additionally, the framework implements hysteresis to prevent sending too many notifications. Thus for the above example, the framework:

  • Sends one notification when the temperature reaches 55 °C

  • Waits until the temperature drops below 54 °C

  • Sends another notification when the temperature rises back to 55 °C

To generate these notifications, the thermal framework sets low trips on the sensors to receive events that the temperature has dropped below the limit.

Hardware Throttling

Each element in a power delivery system includes limitations such as:

  • The amount of current a battery can supply without shutting down

  • The amount of current a regulator can provide before it fails to maintain its output voltage

  • The amount of ripple current an inductor in a switching regulator can tolerate without overheating

These limitations can result in fast transient electrical and thermal events such as:

  • Overcurrent at the battery

  • Voltage drop at the PMIC

  • Temperature spikes

The firmware refers to these events as OC alarms, and triggers clock hardware throttling to handle them.

Impact

Like software throttling, hardware throttling may reduce performance. Because the triggering events are rare and transient in nature, though, the user experience is minimally impacted.

The host OS is not notified of these events, but you can detect the drop in clock rates by using a performance measuring tool that samples the CPU cycle counters. While thermal management in the host OS seeks to control temperature on an ongoing basis, hardware throttling clamps down the clocks to handle events.

Throttle Points and Vector Configuration

The BPMP device tree binary holds the various throttle points and the throttle settings that govern when and how throttling is performed. The soctherm driver in the BPMP firmware handles any interrupts resulting from these events. You can change the throttle points by changing the BPMP device tree.

This table shows the hardware throttling levels:

Hardware throttling

Clock throttled percentage

Heavy

87.5

Medium

75

Light

50

Throttle vectors are optimized for limiting peak current consumption while maximizing performance. To manage peak current consumption, the firmware supports capping the CPU and GPU clocks at three levels (light, medium, and heavy), as described in the device tree bindings. Clock capping prevents the CPU and GPU from drawing more current than their voltage regulators can supply.

For hardware throttling trip temperatures, see the table in Thermal Specifications.

Design Considerations

Designing failsafe measures into Power Management Integrated Circuits (PMICs), or using the battery controller to shut down the device when the events described here occur, results in a bad user experience. Similarly, designing power delivery hardware for worst-case loads results in large and costly components.

Consequently, NVIDIA SoCs are designed for use with power delivery systems that are adequate for common loads. NVIDIA SoCs actively manage their components to avoid exceeding their design limits. When events are transient, the advantage of this approach to power management becomes more compelling.

Hardware Thermal Shutdown

The final failsafe for firmware thermal management is a hardware thermal reset, or thermtrip. If software and hardware throttling are unable to control heat generation in the system, and the software becomes unresponsive, the SoC asserts the reset pin on the PMIC as the hardware shutdown mechanism.

For hardware shutdown limits, see the table in Thermal Specifications.

Thermal Specifications

This table describes the supported cooling states.

Thermal zone

Thermal sensor

Cooling action

Jetson Xavier NX series

Jetson AGX Xavier & Jetson AGX Xavier 64GB

Jetson AGX Xavier Industrial

10W - 30W

10W - 30W [1]

max EDP [2]

20W - 40W [3]

max EDP [4]

CPU-therm

THERMAL_ZONE_PLLX

SW throttling

90.5 °C

90.0 °C

86.0 °C

103.0 °C

96.0 °C

HW throttling

94.5 °C

94.0 °C

90.0 °C

107.0 °C

100.0 °C

SW shutdown

96.0 °C

95.5 °C

91.5 °C

108.5 °C

101.5 °C

HW shutdown

96.5 °C

96.0 °C

92.0 °C

109.0 °C

102.0 °C

GPU-therm

THERMAL_ZONE_GPU

SW throttling

91.5 °C

92.5 °C

88.0 °C

102.0 °C

98.0 °C

HW throttling

95.5 °C

96.5 °C

92.0 °C

106.0 °C

102.0 °C

SW shutdown

97.0 °C

98.0 °C

93.5 °C

107.5 °C

103.5 °C

HW shutdown

97.5 °C

98.5 °C

94.0 °C

108.0 °C

104.0 °C

AUX-therm

THERMAL_ZONE_AUX

SW throttling

90.0 °C

89.0 °C

82.0 °C

100.5 °C

92.5 °C

HW throttling

94.0 °C

93.0 °C

86.0 °C

104.5 °C

96.5 °C

SW shutdown

95.5 °C

94.5 °C

87.5 °C

106.0 °C

98.0 °C

HW shutdown

96.0 °C

95.0 °C

88.0 °C

106.5 °C

98.5 °C

AO-therm

THERMAL_ZONE_AO

HW shutdown

109.0 °C

109.0 °C

109.0 °C

119.0 °C

119.0 °C

PMIC-Die

PMIC thermal sensor

HW shutdown

120.0 °C

120.0 °C

120.0 °C

120.0 °C

120.0 °C

Tboard-tegra

TMP451 local sensor

HW shutdown

-

107.0 °C

107.0 °C

125.0 °C

125.0 °C

Tdiode-tegra

TMP451 external sensor

HW shutdown

-

109.0 °C

109.0 °C

117.0 °C

117.0 °C

[1] Jetson AGX Xavier and Jetson AGX Xavier 64GB use the 10W - 30W thermal specification when it is flashed with jetson-xavier.conf.
[2] Jetson AGX Xavier and Jetson AGX Xavier 64GB use the max EDP thermal specification when it is flashed with jetson-xavier-maxn.conf.
[3] Jetson AGX Xavier Industrial uses the 20W - 40W thermal specification when it is flashed with jetson-agx-xavier-industrial.conf.
[4] Jetson AGX Xavier Industrial uses the max EDP thermal specification when it is flashed with jetson-agx-xavier-industrial-mxn.conf.

Software-Based Power Consumption Modeling

Jetson Xavier modules integrate a three-channel INA3221 power monitor whose information can be read using sysfs nodes. The following table shows the naming convention for sysfs nodes.

Command*

Description

in<n>_label

Sets/gets rail name.

curr<n>_input

Gets rail current in milliamperes.

in<n>_input

Gets rail voltage in millivolts.

curr<n>_crit

Sets/gets rail instantaneous current limit in milliamperes.

curr<n>_max

Sets/gets rail average current limit in milliamperes.

* <n> is a channel number, 1-3.

Note

The INA driver may also present other nodes. Do not modify any INA sysfs node value. Modifying these values can result in damage to the device.

Jetson Xavier NX Series

Jetson Xavier NX series has one INA3221 power monitor at I2C address 0x40. The sysfs nodes to read for rail names, voltage, current, and instantaneous and average current limit are at:

/sys/bus/i2c/drivers/ina3221/7-0040/hwmon/hwmon<x>
(Where <x> is dynamic hwmon index)

The rail names for I2C address 0x40 are:

Channel Name

Description

Channel 1: VDD_IN

Total Module Power

Channel 2: VDD_CPU_GPU_CV

Total power consumed by CPU, GPU and CV cores i.e. DLA and PVA

Channel 3: VDD_SOC

Power consumed by SOC core which supplies to memory subsystem and various engines like nvdec, nvenc, vi, vic, isp etc.

Here are the configurations of TDP average and instantaneous power limits for Jetson Xavier NX modules:

Module

NVP Model Mode

Limits

SOCTHERM_OC PIN

Throttling Level

Jetson Xavier NX

20W

VDD_IN Average Power: 20W

OC3

  • CPU: None

  • GPU: 50%

VDD_IN Instantaneous Power: 25W

OC1

  • CPU: 50%

  • GPU: 50%

Under Voltage: approx. 4.5V

OC2

  • CPU: 87.5%

  • GPU: 87.5%

Jetson Xavier NX

15W and 10W

VDD_IN Average Power: 15W

OC3

  • CPU: None

  • GPU: 50%

VDD_IN Instantaneous Power: 25W

OC1

  • CPU: 50%

  • GPU: 50%

Under Voltage: approx. 4.5V

OC2

  • CPU: 87.5%

  • GPU: 87.5%

Note

The VDD_IN rail overcurrent thresholds for average current are 3  A for 10W and 15W power modes, and 4  A for 20W power modes. For instantaneous current they are 5  A for all power modes. When module current consumption exceeds the configured limits, the INA3221 triggers CPU and GPU hardware clock throttling to prevent shutdown and physical damage. The nvpmodel GUI applet notifies user space processes of overcurrent events. For more details, see the section Voltage and Current Monitor.

Jetson AGX Xavier Series

The Jetson AGX Xavier series modules have two 3-channel INA3221 power monitors at I2C addresses 0x40 and 0x41. The sysfs nodes to read for rail names, voltage, current, and instantaneous and average current limit are at:

/sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>
/sys/bus/i2c/drivers/ina3221/1-0041/hwmon/hwmon<y>
(Where <x> and <y> are dynamic hwmon indexes)

The rail names for I2C address 0x40 are:

Channel Name

Description

Channel 1: GPU

Power consumed by GPU

Channel 2: CPU

Power consumed by CPU

Channel 3: SOC

Power consumed by SOC core which supplies to memory subsystem and various engines like nvdec, nvenc, vi, vic, isp etc.

The rail names for I2C address 0x41 are:

Channel Name

Description

Channel 1: CV

Power consumed by CV cores i.e. DLA and PVA

Channel 2: VDDRQ

Power consumed by DDR core

Channel 3: SYS5V

Power consumed by system 5V rail which supplies to various IOs e.g. HDMI, USB, SDMMC, EMMC etc.

Examples

  • To read INA3221 at 0x40, the channel-1 rail name, enter the command:

    $ cat /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/in1_label
    
  • To read channel-1 voltage and current, enter the commands:

    $ cat /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/in1_input
    $ cat /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/curr1_input
    
  • To read the channel-1 instantaneous current limit, enter the command:

    $ cat /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/curr1_crit
    
  • To set the channel-1 instantaneous current limit, enter the command:

    $ echo <current> > /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/curr1_crit
    
  • To read the channel-1 average current limit, enter the command:

    $ cat /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/curr1_max
    
  • To set the channel-1 average current limit, enter the command:

    $ echo <current> > /sys/bus/i2c/drivers/ina3221/1-0040/hwmon/hwmon<x>/curr1_max
    

    Where <current> is the current limit to be set for the rail, in milliamperes.

Overcurrent Event Status

The following OC events types exist in the Jetson Xavier series:

  • Instantaneous Overcurrent

  • Under Voltage

  • Average Overcurrent

The following sysfs nodes can be used to learn the number of OC events occurred:

$ grep "" /sys/class/hwmon/hwmon<x>/oc*_irq_cnt