The Nvidia Jetson Nano - a powerful alternative to the likes of RPi/Movidius bundle when you are looking for a powerful edge device, which can cope with some machine learning/deep learning tasks. This page provides some details about hardware and software related topics - stuff, which had been useful to do the first steps with the device.
Note: Jetpack 4.4 performance issues: It has been reported, that the performance on Nano drops significantly (50% or more) when updating to Jetpack 4.4. There are related posts like this , which hint not to use Jetpack 4.4 for now.
The Nvidia box, which comes woith the device is pretty much useless for serious work. Just throw it away and start the beefing up right away - or at least prepare for it.
To get a nice to handle package a 3D printed case has been used. The needed case package can be found on the Thingyverse site. The case, done by Thingyverse user ecoiras fits snuggly and is very well done. Only the connector side is a bit delicate and breaks easily.
Well, if you print it with a cheap printer and just one color the case looks like this:
A look into the inside:
In the meantime a simple acrylic commercial case is in use. The product can be found on Amazon.
The Jetson Nano can run in two different power modes. Per default the board is running in the "10 Watt" mode (also known as Mode 0) but there is also the optional/fallback "5 Watt" mode. The 10 Watt mode has a catch: The Jetson module itself already consumes 10 Watt when all cores are up and running. There is no more reserve available if you are running the board via the Micro-USB connector with 5V and 2A max. The module does not get what it wants since the Jetson basis board and all connected USB devices drain some power too.
During the first test runs the board was still fine (mouse, keyboard, Webcam) and it ran also nicely even after adding the fan. Still it is highly recommended to switch to a 5V 4A power supply with the Barrel Jack connector (5V inside). The shown power supply (LEICKE Power Supply 5V 4A | 20W) works fine with the device - you may find others from the usual sources.
To activate the Barrel Jack power supply a Jumper needs to be set. The related jumper (not part of the Jetson Nano board delivery so remember where you last saw jumpers laying around) is located next to the camera interface:
Now everything runs inside the specs and the power hungry DL inference can run safely.
Show current power mode:
sudo nvpmodel -q
The output migh look like this:
NV Power Mode: MAXN
0
To set mode 0 (10W):
sudo nvpmodel -m 0
To set mode 1 (5W):
sudo nvpmodel -m 1
It is possible to define your own nvpmodel - the respective configuration file can be found here:
nano /etc/nvpmodel/nvpmodel_t210_jetson-nano.conf
The configuration allows one to change clock rates and many other details (to be covered later). But the Power_Model Definitions sections is revealing:
###########################
# #
# POWER_MODEL DEFINITIONS #
# #
###########################
# MAXN is the NONE power model to release all constraints
< POWER_MODEL ID=0 NAME=MAXN >
CPU_ONLINE CORE_0 1
CPU_ONLINE CORE_1 1
CPU_ONLINE CORE_2 1
CPU_ONLINE CORE_3 1
CPU_A57 MIN_FREQ 0
CPU_A57 MAX_FREQ -1
GPU_POWER_CONTROL_ENABLE GPU_PWR_CNTL_EN on
GPU MIN_FREQ 0
GPU MAX_FREQ -1
GPU_POWER_CONTROL_DISABLE GPU_PWR_CNTL_DIS auto
EMC MAX_FREQ 0
< POWER_MODEL ID=1 NAME=5W >
CPU_ONLINE CORE_0 1
CPU_ONLINE CORE_1 1
CPU_ONLINE CORE_2 0
CPU_ONLINE CORE_3 0
CPU_A57 MIN_FREQ 0
CPU_A57 MAX_FREQ 918000
GPU_POWER_CONTROL_ENABLE GPU_PWR_CNTL_EN on
GPU MIN_FREQ 0
GPU MAX_FREQ 640000000
GPU_POWER_CONTROL_DISABLE GPU_PWR_CNTL_DIS auto
EMC MAX_FREQ 1600000000
As can be seen the power consumption is controlled in a very easy manner: Mode 1 is just using half of the available CPU cores running at reduced speed. Thus, if you want to use all of the available oooomph from the Nano, always run it in Mode 0 and power the board accordingly with the Barrel Jack connector and a decent 4A supply.
The configuration defines some more parameters including the minimum and maximum frequencies per core. These settings are needed to describe the Dynamic Voltage and Frequency Scaling (DVFS) behavior. DVFS is a mechanism to dynamically crank up the CPU depending on the current system load.
A call to
sudo jetson_clocks --show
reveals the current settings:
SOC family:tegra210 Machine:jetson-nano
Online CPUs: 0-3
CPU Cluster Switching: Disabled
cpu0: Online=1 Governor=schedutil MinFreq=102000 MaxFreq=1428000 CurrentFreq=1428000 IdleStates: WFI=1 c7=1
cpu1: Online=1 Governor=schedutil MinFreq=102000 MaxFreq=1428000 CurrentFreq=1428000 IdleStates: WFI=1 c7=1
cpu2: Online=1 Governor=schedutil MinFreq=102000 MaxFreq=1428000 CurrentFreq=1428000 IdleStates: WFI=1 c7=1
cpu3: Online=1 Governor=schedutil MinFreq=102000 MaxFreq=1428000 CurrentFreq=1428000 IdleStates: WFI=1 c7=1
GPU MinFreq=76800000 MaxFreq=921600000 CurrentFreq=153600000
EMC MinFreq=204000000 MaxFreq=1600000000 CurrentFreq=1600000000 FreqOverride=0
Fan: speed=0
NV Power Mode: MAXN
Now use the next command to save the current values into a *.conf *file:
sudo jetson_clocks --store oldClocks.conf
The content shows how the initial configuration from
/etc/nvpmodel/nvpmodel_t210_jetson-nano.conf turns into the
related and current system device settings:
/sys/devices/system/cpu/cpu0/online:1
/sys/devices/system/cpu/cpu1/online:1
/sys/devices/system/cpu/cpu2/online:1
/sys/devices/system/cpu/cpu3/online:1
/sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq:102000
/sys/devices/system/cpu/cpu1/cpufreq/scaling_min_freq:102000
/sys/devices/system/cpu/cpu2/cpufreq/scaling_min_freq:102000
/sys/devices/system/cpu/cpu3/cpufreq/scaling_min_freq:102000
/sys/devices/system/cpu/cpu0/cpuidle/state0/disable:0
/sys/devices/system/cpu/cpu0/cpuidle/state1/disable:0
/sys/devices/system/cpu/cpu1/cpuidle/state0/disable:0
/sys/devices/system/cpu/cpu1/cpuidle/state1/disable:0
/sys/devices/system/cpu/cpu2/cpuidle/state0/disable:0
/sys/devices/system/cpu/cpu2/cpuidle/state1/disable:0
/sys/devices/system/cpu/cpu3/cpuidle/state0/disable:0
/sys/devices/system/cpu/cpu3/cpuidle/state1/disable:0
/sys/devices/57000000.gpu/devfreq/57000000.gpu/min_freq:76800000
/sys/devices/57000000.gpu/railgate_enable:1
/sys/kernel/debug/clk/override.emc/clk_state:0
/sys/devices/pwm-fan/target_pwm:0
/sys/devices/pwm-fan/temp_control:1
Note: When idling, the temperature reaches about 50 degrees Celsius when the Nano based on the default settings. Your mileage may vary depending on the current environmental temperature. It may be a good idea to watch the temp when playing around. The needed script/info about temperatures can be found in the "Useful Scripts" chapter.
Save the configuration file. Now it is save to activate the maximum values right away to check it out:
sudo jetson_clocks
The updated configuration file looks like this:
sudo jetson_clocks --store oldClocks_new.conf
and check out the new values:
/sys/devices/system/cpu/cpu0/online:1
/sys/devices/system/cpu/cpu1/online:1
/sys/devices/system/cpu/cpu2/online:1
/sys/devices/system/cpu/cpu3/online:1
/sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq:1428000
/sys/devices/system/cpu/cpu1/cpufreq/scaling_min_freq:1428000
/sys/devices/system/cpu/cpu2/cpufreq/scaling_min_freq:1428000
/sys/devices/system/cpu/cpu3/cpufreq/scaling_min_freq:1428000
/sys/devices/system/cpu/cpu0/cpuidle/state0/disable:1
/sys/devices/system/cpu/cpu0/cpuidle/state1/disable:1
/sys/devices/system/cpu/cpu1/cpuidle/state0/disable:1
/sys/devices/system/cpu/cpu1/cpuidle/state1/disable:1
/sys/devices/system/cpu/cpu2/cpuidle/state0/disable:1
/sys/devices/system/cpu/cpu2/cpuidle/state1/disable:1
/sys/devices/system/cpu/cpu3/cpuidle/state0/disable:1
/sys/devices/system/cpu/cpu3/cpuidle/state1/disable:1
/sys/devices/57000000.gpu/devfreq/57000000.gpu/min_freq:921600000
/sys/devices/57000000.gpu/railgate_enable:0
/sys/kernel/debug/clk/override.emc/clk_state:1
/sys/devices/pwm-fan/target_pwm:255
/sys/devices/pwm-fan/temp_control:0
After calling the mighty jetson_clocks command, the system runs at full speed including the fan (if connected). As a result the temperature drops to 36 degrees when idling.
To switch back to the default settings:
sudo jetson_clocks --restore oldClocks.conf
With "normal" displays there where no real problems occurring with the exception of the wrong display resolution. When connecting to a Ilyama XUB3493WQSU wide screen the dashboard did not show up.
To check the supported display resolutions use the following command:
xrandr
To set a specific resolution use the following (example) command:
xrandr --output HDMI-0 --mode "2560x1440"
A good explanation of the xrandr command cen be found on this blog. To persist the setting, the command can be added to the .xprofile file or by using the xorg.conf alternative. More information can be found here .
When running the first real inference tasks the CPU temperature easily jumped over 60 degrees and it seemed to have an negative impact on system stability as the nano occasionally went into freezing mode from time to time. The first hardware option added has been the fan. As recommended by others in forums etc. the "Noctua NF-A4x20 5V PWM" has been chosen. This fan is a bit on the expensive side, but very well built and very silent. The fan comes with a lot of additional material, which is of no use in this environment. To fix the fan two slim zip ties have been used.
The fan control is based on the current temperature but it is also possible to activate or de-activate the fan via a command.
Activating the fan:
sudo sh -c 'echo 255 > /sys/devices/pwm-fan/target_pwm'
De-Activating the Fan:
sudo sh -c 'echo 0 > /sys/devices/pwm-fan/target_pwm'
Since the fan activation on the board is temperature controlled there is usually no need to explicitly activate/deactivate it though.
The SD card is the primary non-volatile memory system for the Nano. The SD card stores the complete operating system, all data models (if needed) and of cause all applications and the respective code you may run. Thus it is always a good idea to use a 64 GByte SD card, which will give you enough head room for everything.
But if you need additional memory for even more training/inference data, faster general data access or if you want to add some decent swap space an external SSD is of great use. For the work to be done a SanDisk 1 TB SSD (SanDisk SSD PLUS 1TB Sata III 2,5 Inch internal-SSD, 535MB/Sek) has been chosen. Check out the options - the price of this device was moving between 88 and 140+ E>uros in one week.
To be able to connect it to the USB 3.0 slot a adapter is needed (CSL - USB 3.0 SSD SATA Adapter or comparable):
To manage the file system you can either use the already installed parted command or use the more intuitive graphical UI for parted. This UI comes with the gparted application. To install the gparted application use the next command:
sudo apt-get install gparted
With just the initial SD card inserted the information may look like this:
Now its time to define the partition info for the SSD. First a partition table (MSDOS) has to be set for the complete device. Then its up to you to define the partition details.
In the following example 8 GByte have been dedicated as linux swap and the reminder of the available 1 TB are mapped to the /mnt/data folder. This folder needs to be present in the root file system to allow for a later file system mount. Check the chosen settings and select the green arrow on the gparted main menu. This will start the partitioning - all current data on the SSD will be gone after this step.
Now its time to define the mapping for two new partitions by editing the fstab file:
sudo nano /etc/fstab
Change the fstab file along the lines of this example:
/dev/root / ext4 defaults 0 1
/dev/sda1 swap swap defaults 0 0
/dev/sda2 /mnt/data ext4 defaults 0 2
Now you can already activate the swap with this command:
sudo swapon -a
A quick check in the system monitor will reveal the new swap swap being available.
During the next boot the /mn/data mount point is mounted with the respective SSD partition - in our case /dev/sda2.
Note: It is currently not possible to boot from SSD directly.
Note: It is not recommended to use the SD card as swap space. While possible you would wear out the memory region defined by the static swap file (something like "/mnt/swapfile") pretty fast.
Jetson Stats - Simple package to monitoring and control your NVIDIA Jetson [Nano, Xavier, TX2i, TX2, TX1]
The current state of important temperature values can be retrieved via this command:
cat /sys/devices/virtual/thermal/thermal_zone*/temp
Result:
50000
41500
40500
40500
100000
41250
The explanation of the given distinct values are explained via this command:
cat /sys/devices/virtual/thermal/thermal_zone*/type
A good general ressource is the Jetson/Thermal blog from elinux.org. This blog provides also the code for a simple "showTemp.pl" Perl script, which will show the current temperature. Can come in very handy when running intense use cases.
The following command displays some important system parameters like various temperature values, CPU core loads or currently available memory:
sudo tegrastats
While the file /etc/rc.local has gone with Ubuntu 18.04 it is still possible to add this file manually. If present the contents will be read and used as the last step of the system boot phase.
$ cat /etc/rc.local
#!/bin/bash
sleep 10
sudo /usr/bin/jetson_clocks
sudo sh -c 'echo 255 > /sys/devices/pwm-fan/target_pwm'
There is a blog on the Nvidia Developer Forum covering installing PyTorch on the Nano
If you want to buy some of the products covered, i have listed the links below.
SanDisk SSD PLUS 1TB Sata III 2,5 Zoll Interne-SSD, bis zu 535MB/Sek
Noctua NF-A4x20 5V PWM, Leiser Premium-Lüfter, 4-Pin, 5V Version (40x20mm Braun)
SanDisk SSD PLUS 1TB Internal SSD - SATA III 6 Gb/s, 2.5"/7mm - SDSSDA-1T00-G26
Noctua NF-A4x20 PWM, Premium Quiet Fan, 4-Pin (40x20mm, Brown)