Nvidia Jetson

The Jetson Family

Nvidia's Jetson family of products are compact Tegra computers, which feature a large number of parallel computation units that are suitable for hardware-accelerated machine learning applications. The various models differ in their computational performance.

JetPack

JetPack is Nvidia's name for a software bundle that contains development software and a modified version of Ubuntu Linux with a custom kernel for the Tegra hardware, called Linux For Tegra or L4T.

• JetPack download page

Figuring out the JetPack Version

The JetPack Archive shows a list of JetPack download packages with corresponding versions of L4T (Linux for Tegra).

In a running L4T installation, the version can be looked up via the Debian package manager utility:

dpkg-query --show nvidia-l4t-core

Some JetPack and L4T versions are listed below for convenience.

L4T	JetPack	release date	Ubuntu base	Remark
35.1	5.0.2	2022-08-15	20.04	5.x drops support for Jetson Nano
32.7.1	4.6.1	2022-02-23	18.04
32.5	4.5.0	2021-01-22	18.04
32.4.3	4.4.0	2020-07-06	18.04
32.2.0	4.2.1	2019-07-16	18.04

Assigning a Fixed Ethernet IP Address

You can either use the old ifconfig tool

ifconfig eth0 192.168.0.10 netmask 255.255.255.0 up

or the newer ip tool (the example below is incomplete)

ip addr show
ip link set eth0 up

To make the IP address permanent, the easy way is to open the System Settings window from the Ubuntu GUI. Then

• open the Network settings,
• select the wired connection from the left hand list,
• click on the "Options..." button in the bottom right corner,
• open the "IPv4 Settings" tab,
• change the value of the "Method:" field to "Manual",
• enter the desired address (192.168.0.10) into the "Addresses" list.

Setting up WiFi

First, make sure that the WLAN hardware (maybe a USB stick) is operational. Check the availability of the wlan0 interface among all network interfaces:

ifconfig

Also, make sure that wpasupplicant is installed.

sudo apt-get install wpasupplicant

Now edit the file /etc/wpa_supplicant.conf to replace the "ssid" and "psk" values in the network example that starts with the line # Simple case: WPA-PSK:

network={
    ssid="ssid_name"
    psk="password"
    priority=5
}

where ssid_name and password need to be replaced with the WiFi name (SSID) and the password of your router, respectively.
Finally, connect to the router:

sudo killall wpa_supplicant
sudo wpa_supplicant -B -i wlan0 -c /etc/wpa_supplicant.conf -D wext
sudo dhclient wlan0

Setting up SSH

The developer kit features HDMI and DisplayPort outputs for connecting to a display. In a robot, the control interface for the Jetson Nano will be a terminal session via SSH. To set up the SSH service on the Nano first make sure that the ssh package is installed:

sudo apt install openssh-client openssh-server

You can now start the SSH service like this:

sudo service ssh start

and stop it like this:

sudo service ssh stop

To make the SSH service start automatically at each boot, type

sudo systemctl enable ssh
sudo systemctl enable ssh.service
sudo systemctl start ssh

To prevent the system from starting the SSH service at each boot, type

sudo systemctl stop ssh
sudo systemctl disable ssh
sudo systemctl disable ssh.service

SSH Login Over WiFi Connection

The Jetson Nano and the Jetson Xavier developer kit boards come without WiFi connectivity. You can use a USB WiFi adaptor stick (e.g., Edimax ) to add WiFi connectivity. On Ubuntu, you will need to give permission to the system to make the WiFi connection available to all other users, which includes the SSH daemon, in order to enable remote logins over WiFi without requiring an active local login session (via connected peripherals or Ethernet connection).
To give permission,

• open the "System Settings" application,
• go to "Network",
• make sure "Wireless" is selected on the left panel,
• on the right side panel, next to the SSID of the WLAN server, click on the arrow icon,
• click on the "Settings..." button on the bottom right corner,
• in the window that opens, go to the tab "General",
• check the "All users may connect to this network" option

Camera

Cameras can be connected to Jetson Nano over USB or over the CSI connector. In case of a CSI camera, the image sensor must be a Sony IMX219, such as the Raspberry Pi Camera v2.
Use the Video4Linux command-line tool v4l2-ctl to list all video inputs that are available to the system and what type of image frames they provide.

sudo apt install v4l-utils
v4l2-ctl --list-devices
v4l2-ctl --list-formats

You can view the video stream from the CSI camera using GStreamer via

gst-launch-1.0 nvarguscamerasrc ! 'video/x-raw(memory:NVMM), width=1920, height=1080, format=NV12, framerate=(fraction)30/1' ! nvoverlaysink

for a full-screen view, or

gst-launch-1.0 nvarguscamerasrc ! 'video/x-raw(memory:NVMM), width=1920, height=1080, framerate=30/1' ! nvvidconv flip-method=0 ! nvegltransform ! nveglglessink -e

for an OpenGL window view.

nvarguscamerasrc is a Gstreamer plugin for Nvidia's proprietary Argus Camera API for Linux on Tegra devices, and nveglglessink is a better-performing OpenGL sink than glimagesink. Curiously, nvarguscamerasrc will show up in the GStreamer element list when searching for elements of type video but not when searching for source/video.

gst-inspect-1.0 -t video

Streaming Images Over The Network

To stream the camera images over the network via UDP, run the following script on the Jetson

#!/bin/bash
STREAM_RECEIVER_ADDRESS=$1
STREAM_RECEIVER_PORT=$2
GST_LAUNCH="gst-launch-1.0 -v"
GST_DEBUG="--gst-debug=3"
INPUT="nvarguscamerasrc ! 'video/x-raw(memory:NVMM),width=1280,height=720,framerate=30/1,format=(string)NV12' ! nvvidconv ! 'video/x-raw(memory:NVMM),format=(string)I420'"

# Stream via RTP over UDP
STREAM_OUTPUT="omxh264enc bitrate=8000000 ! 'video/x-h264, stream-format=byte-stream' ! h264parse ! rtph264pay config-interval=10 pt=96 ! udpsink host=<host address> port=<port number>"

# Show in local GL window
VIEW_OUTPUT="nvegltransform ! nveglglessink -e"

if [ $# == 0 ]; then
    eval $GST_LAUNCH $INPUT ! $VIEW_OUTPUT
elif [ $# == 2 ]; then
    eval $GST_LAUNCH $INPUT ! $STREAM_OUTPUT
fi

where the client IP address and the client port number need to be supplied as arguments 1 and 2, respectively.

Run the following script on the client machine to receive and display the video stream in a GStreamer pipeline.

UNPACK_RTP_H264_PAYLOAD="'application/x-rtp, media=(string)video, clock-rate=(int)90000, encoding-name=H264, payload=96' ! rtph264depay ! avdec_h264"
GST_UDPSRC="gst-launch-1.0 -v --gst-debug=3 udpsrc port=<port number>"
VIDEO_SINK="videoconvert ! autovideosink"

eval $GST_UDPSRC ! $UNPACK_RTP_H264_PAYLOAD ! $VIDEO_SINK

where you need to replace <port number> with the same port number used by the sender. Check the receiver's firewall settings to make sure the port is not blocked! On macOS you can do so by opening the Firewall settings and first add Terminal to the list of applications for which incoming connections are allowed. Upon launching the script, you will be prompted to allow incoming connections for gst-launch-1.0 as well.

The order in which the Gstreamer pipelines are launched is irrelevant because of the connectionless data transmission via UDP. The receiver will be continuously waiting for incoming UDP packets, and the sender will be sending out packets even if there is no receiver.

Resources

• example GStreamer pipelines (RidgeRun)

Power Modes

Use the nvpmodel command line tool for selecting or configuring the performance level of a Jetson device. You can query the current mode via

sudo nvpmodel -q --verbose

and switch between modes with the "-m" option like this

sudo nvpmodel -m <mode_number>

where <mode_number> is one of the mode numbers (e.g., 0 or 1) which the tool reads from the default configuration file "/etc/nvpmodel.conf" and which you can edit to create a custom mode:

sudo vim /etc/nvpmodel.conf

Disabling USB Autosuspend

If you are operating your Jetson with an attached USB mass storage device you will need to disable USB autosuspend. Otherwise the device will become unavailable after a while, and your applications will report an I/O error.

Edit the file /boot/extlinux/extlinux.conf by adding the following boot argument to the line starting with APPEND:

usbcore.autosuspend=-1

This will disable autosuspend upon the next reboot.

To check that the autosuspend mode changed from active (2) to disabled (-1), enter

cat /sys/module/usbcore/parameters/autosuspend

GPIO Programming

Jetson Nano Pinout

Based on the JetsonHacks pinout table.

Sysfs GPIO	Name	Pin		Pin	Name	Sysfs GPIO
	3.3 V Power out	1		2	5.0 V Power in
	I2C_2_SDA	3		4	5.0 V Power in
	I2C_2_SCL	5		6	GND
gpio216	AUDIO_MCLK	7		8	UART_2_TX
	GND	9		10	UART_2_RX
gpio50	UART_2_RTS	11		12	I2S_4_SCLK	gpio79
gpio14	SPI_2_SCK	13		14	GND
gpio194	LCD_TE	15		16	SPI_2_CS1	gpio232
	3.3 V	17		18	SPI_2_CS0	gpio15
gpio16	SPI_1_COPI	19		20	GND
gpio17	SPI_1_CIPO	21		22	SPI_2_CIPO	gpio13
gpio18	SPI_1_SCK	23		24	SPI_1_CS0	gpio19
	GND	25		26	SPI_1_CS1	gpio20
	I2C_1_SDA	27		28	I2C_1_SCL
gpio149	CAM_AF_EN	29		30	GND
gpio200	GPIO_PZ0	31		32	LCD_BL_PWM	gpio168
gpio38	GPIO_PE6	33		34	GND
gpio76	I2S_4_LRCK	35		36	UART_2_CTS	gpio51
gpio12	SPI_2_COPI	37		38	I2S_4_SDIN	gpio77
	GND	39		40	I2S_4_SDOUT	gpio78

Using SYSFS

The old, deprecated way of GPIO programming on Linux is to write to the sysfs virtual files that are managed by the kernel. For GPIO, the Linux kernel provides the /sys/class/gpio virtual directory. GPIO commands, like creating an output pin and writing a value, are performed by writing to or reading from specific files in that directory.

Since sysfs is a filesystem-based interface, GPIO state can be easily manipulated with a shell script. Here is an example bash code for blinking an LED connected to GPIO pin 11:

echo 50 > /sys/class/gpio/export 
sleep 0.1
echo out > /sys/class/gpio/gpio50/direction
sleep 0.1
for n in {0..2}
    do
      echo 1 > /sys/class/gpio/gpio50/value
      sleep 1
      echo 0 > /sys/class/gpio/gpio50/value
      sleep 1
    done
echo 50 > /sys/class/gpio/unexport

Using Libgpiod

The best way for programming GPIO without having to mess with the inefficient sysfs, is to use the C library libgpiod or a binding to another programming language. First, make sure libgpiod, the C header, and the related command-line tools are installed

sudo apt install gpiod libgpiod-dev

gpiodetect gives us information about the available GPIO banks. Typing

sudo gpiodetect

returns

gpiochip0 [tegra-gpio] (256 lines)
gpiochip1 [max77620-gpio] (8 lines)

where gpiochip0 is the GPIO bank exposed by the J41 header.

Drilling down into GPIO bank 0 with gpioinfo

gpioinfo gpiochip0

we can see the pin functions of J41 GPIO bank:

gpiochip0 - 256 lines:
        line   0:      unnamed       unused   input  active-high
        line   1:      unnamed       unused   input  active-high
        line   2:      unnamed  "pcie_wake"   input  active-high [used]
        line   3:      unnamed       unused   input  active-high
        line   4:      unnamed       unused   input  active-high
        line   5:      unnamed       unused   input  active-high
        line   6:      unnamed "vdd-usb-hub-en" output active-high [used]
        line   7:      unnamed       unused   input  active-high
        line   8:      unnamed       unused   input  active-high
        line   9:      unnamed       unused   input  active-high
        line  10:      unnamed       unused   input  active-high
        line  11:      unnamed       unused   input  active-high
        line  12:  "SPI1_MOSI"       unused   input  active-high
        line  13:  "SPI1_MISO"       unused   input  active-high
        line  14:   "SPI1_SCK"       unused   input  active-high
        line  15:   "SPI1_CS0"       unused   input  active-high
.
.
.
        line 252:      unnamed       unused   input  active-high
        line 253:      unnamed       unused   input  active-high
        line 254:      unnamed       unused   input  active-high
        line 255:      unnamed       unused   input  active-high    

Using the gpioset shell command we can set the output level of a GPIO pin:

sudo gpioset gpiochip0 50=1

If the LED does not light up, check if line 50 is still being used by sysfs.

Back to programming the GPIO in C++. Create the file libgpiod_test.cpp containing the following C++ code:

#include <gpiod.h>
#include <cstdlib>
#include <thread>
#include <chrono>

int main(int numarg, char** args)
{
    gpiod_chip* const gpio0 = gpiod_chip_open_by_name("gpiochip0");
    if (gpio0 == NULL) {
        return EXIT_FAILURE;
    }
    gpiod_line* const line50 = gpiod_chip_get_line(gpio0, 50);
    if (line50 == NULL) {
        gpiod_chip_close(gpio0);
        return EXIT_FAILURE;
    }
    if (gpiod_line_request_output(line50, "LED", 0) != 0) {
        gpiod_line_release(line50);
        gpiod_chip_close(gpio0);
        return EXIT_FAILURE;
    }
    for (int i = 1; i <= 6; i++)
    {
        gpiod_line_set_value(line50, i%2);
        std::this_thread::sleep_for(std::chrono::milliseconds(1000));
    }
    gpiod_line_release(line50);
    gpiod_chip_close(gpio0);
    return EXIT_SUCCESS;
}

Compile and run.

g++ -o libgpiod_test libgpiod_test.cpp -lgpiod
sudo ./libgpiod_test

Debug

To view the state of all the available GPIO pins type

sudo cat /sys/kernel/debug/tegra_gpio

Jetson Model Specifics

Jetson Nano

The Nano is the Jetson with the smallest footprint and lowest performance. It features 128 compute units based on the Maxwell architecture. The developer kit comes with host circuit board that provides connection ports like USB, Ethernet, HDMI and DisplayPort.

• Jetson Nano Developer Kit User Guide

Power Supply and OS Installation

The Nano Developer Kit requires a power supply that can deliver 5 V and at least 2 A, either over the micro USB port, or via a standard 5.5/2.1mm barrel power jack, which requires bridging the J48 header pins with a jumper. When operated via the power jack, the micro USB port can be used to install the Nano operating system from an external computer. Otherwise, the OS image needs to be written to a micro-SD card that is inserted into the card slot of the developer kit.

Preconfigured Power Modes

• mode 0 (MAXN)
  the default mode with maximum performance and up to 10 W power consumption
• mode 1 (5 W)
  low-power mode where only two out of four CPU cores are working and the GPU clock frequency is constrained to keep power consumption under 5 W.

	MAXN	5 W
power budget	10 Watts	5 Watts
CPUs in use	4	2
CPU max freq.	1479 MHz	918 MHz
GPU max freq.	921 MHz	640 MHz