Flashing and CUDA installation using SDK Manager Installation

Hi All,

I have some queries going forward using SDK manager to flash my Jetson device which sits on a customized carrier board.
Also I want to install CUDA and other deep learning packages related to it using SDK manager.
I have few queries related to these two. Please help me in clarifying these.

  1. in the below links:
    The Extra Configuration File — sdk-manager 2.0.0 documentation
    Install Jetson Software with SDK Manager — sdk-manager 2.0.0 documentation

It explains method of flashing on a customized carrier board. But I did not see any steps, from where my extra configuration file will pick my BSP built Image, DTB files etc for flashing.
Could some one clarify me about this?

  1. Also wanted to know, is there any complete thorough uninstalling the software components on the target option from SDK manager ? or should we manually uninstall one by one using"apt --purge -remove… " commands on the target itself.

  2. Does the command to force the unit to “force recovery mode”

"sudo reboot –force forced-recovery "

still works for Jetpack 5.1.2 on Jetson AGX Xavier Industrial unit. please confirm.

  1. The document is created by other teams, which we do not verify, and are not familiar with.
    We still recommend flashing custom carrier boards with flashing scripts.
  2. NO. Please remove them with apt on the device.
  3. YES.

Thanks for the updates.

I have one more query, I want to use ethernet for installation of software library on the target fr SDK manager.

We need to run some ssh -server installation command on the target before this to work.
Any idea what is the exact command that we need to run?
Pls let us know.

I don’t really know what you are talking about.
SSH server is by default installed, and there is nothing you need to do for it to work.

Ok fine.

One more query,
Do I really need the jetson device connected over USB 3.0 flashing port cable with forced recovery mode for it be detected in the SDK manager?

Or

I can manage with just ethernet connection for flashing the software libraries from SDK manager.

Why I am asking in this question is, for USB flashing port enabling we need to open up our packed box of the unit which is over head.

If we can manage installation of software components with ethernet, then it easy for us.

USB flashing cable is needed, but not force recovery mode.
Ethernet cable will not work here.

Ok.you mean Force recovery mode is needed if I want to reflash the image?

For installing just the software cuda etc library packages recovery mode not needed?

Have one more query. I have the SOM module on the customized carrier board, last week I observed, the SDK manager was not detecting my Jetson AGx xavier industrial device. Any idea why?

Any work around method for fix this board detection issue?

Ok.But they SDK manager was listing ethernet also as one of the options in the list box.

That’s right.
For installing JetPack SDK, your device has to be running in the Linux environment, and able to communicate with the host PC via SSH.

I cannot tell with such limited information.
If you just want to use an Ethernet cable instead of a USB flashing cable, then SDK Manager is not really needed.
Just SSH into the device and run these commands:

sudo apt update
sudo apt install nvidia-jetpack

1 Like

Ok fine. SSH means its over ethernet. We should configure my Jetson device and Host PC with IP address manually and then only they can interact via SSH right. Pls confirm.

Its simple. My jetson device( which already was flashed with my customized BSP kernel, DTB files etc through script method) was connected to my host PC via a USB 3.0 flash enabled cable. But in the SDK manager it was saying “board not detected” as shown below:

image

Is there any need that I should place my SOM module on the developer kit carrier board to make SDK manager detected my board?

  1. What and all does this nvidia-jetpack command install actually?
  2. Basically I wanted to install all supported the software components( related to Jetpack 5.1.2) on my AGX Xavier Industrial SOM as shown below:

JetPack 5.1.2 Components
CUDA 11.4.19
** TensorRT 8.5.2**
** cuDNN 8.6.0**
** VPI 2.3**
** OpenCV 4.5.4**
** Vulkan 1.3**
** Nsight Systems 2022.5**
** Nsight Graphics 2022.6**
** Nsight DLD/Compute 2022.2**

I want to install pytorch, Gstreamer etc which are not listed above.

  1. Instead of doing SSH to the device from my Host PC, cannot we directly execute
    sudo apt update
    sudo apt install nvidia-jetpack

on my Jetson device so that it installs all software library/packages on it

You don’t need to configure anything.
Just make sure the USB flashing cable is plugged in, and you can ping the device on your host PC on 192.168.55.1.

Everything listed here.

Of course you can.

Will try this and let you know.

Here it lists many things including Jetson Linux, which I dont want:

I want only the below software components to be installed, it should not touch my existing /already flashed kernel image/DTB customized BSP on the device.

CUDA 11.4.19
** TensorRT 8.5.2**
** cuDNN 8.6.0**
** VPI 2.3**
** OpenCV 4.5.4**
** Vulkan 1.3**
** Nsight Systems 2022.5**
** Nsight Graphics 2022.6**
** Nsight DLD/Compute 2022.2**

It will not…

I want Jetpack 5.1.2 compatible software components to be installed.
By default with this command:
sudo apt install nvidia-jetpack

might install the latest Jetpack 6.0 right?

Any modification or parameter passing required for the above command, so that it installs software libraries/components for Jetpack 5.1.2.

Please clarify.

NO.
The repo should by default points to the one for 5.1.2, if you are using L4T 35.4.1.
It’s just as simple as it is, and I don’t really want to reiterate it so many times.

Sorry. If I am annoying. Thanks a lot for your updates.

As this is first time I am trying out these stuffs on my production unit, I am getting more information to understand things fully before experimenting anything new.

I have one more query related to memory constraint:

As by default my image/dtb etc is flashed on the root file system on internal default eMMC memory, which has not enough space. Total space of root file system was 24 GB, in that its already used 23 GB. [As I was trying out docker/container method of CUDA package installation and did not work out. 13 GB got used by this docker/container image installation]

I have dedicated NVMM memory of 240 GB free space. I want this Jetpack software components to be installed on the mounted NVMM. Please let me know, how to point my installation to happen on that NVMM mount point?

If not possible, then I want to know, if we can expland the root file system memory 24 GB on my internel eMMC memory to additional extra 33 GB( which is there as “none” residing on my 64 GB capacity eMMC memory)

That is not possible.
Please re-flash the device so it uses NVMe as the rootfs.

I don’t know what you are talking about here.

Please let me know which command we need to use.
I was using this below command to flash normally on default eMMC.

Any idea how much size the Jetpack would be approximately?

$ sudo ./flash.sh jetson-agx-xavier-industrial mmcblk0p1

Here should I change the “mmcblck0p1” to my “NVMM mount point name”
as shown below:

$ sudo ./flash.sh jetson-agx-xavier-industrial “<nvmm_mount_point_name>”

will this do the job?

Sorry for the confusion. I was asking is there any method to enlarge the existing parition of “/” from 24 GB to more memory. anyways leave it, if it not understandable. Thanks.

sudo ./tools/kernel_flash/l4t_initrd_flash.sh --external-device nvme0n1p1
-c tools/kernel_flash/flash_l4t_external.xml
–showlogs --network usb0 jetson-agx-xavier-industrial internal

1 Like

Thanks a lot. I was able to install Jetpack perfectly on my Jetson device. Even a sample CUDA program is executing fine.

Next is , the Jetpack does not install PyTorch. Could you please point me to correct path/ command to install Pytorch compatible with Jetpack 5.1.2

There has been a ton of posts on the Internet teaching you how to install PyTorch on Jetson, and there is no need for another post asking for it.

Googling PyTorch Jetson gives you:

1 Like