Jetson Thor: flashing / USB recovery (RCM) from a non-x86 embedded host (ARM or RISC-V)

Hi,

We are designing a custom carrier board for Jetson Thor. In the final system no x86 host is available. If BCM boot slots become corrupted, the module would have to be recovered over USB 2.0 (Force Recovery Mode) by an embedded host on the same board, running on ARM or RISC-V.

As far as we can see, SDK Manager and the flashing tools in Jetson Linux (l4t_initrd_flash.sh, tegrarcm_v2, etc.) only run on x86_64. The RCM protocol used to communicate with the BootROM is not publicly documented.

Our questions:

Is there a supported way to flash or recover Jetson Thor from an ARM64 or RISC-V host?
Can NVIDIA provide ARM64 builds of the flashing tools (tegrarcm_v2 and related), or a host-side flashing library suitable for an embedded system?
Is the RCM protocol specification available, under NDA if necessary?
Is there any supported method to reprogram the module QSPI other than USB recovery mode?

We are happy to continue the discussion privately if needed.

Thanks,
Federico Bua

You could install sdkmanger on Thor and see if it works to flash another Thor. sdkmanager works from command line.

cat /proc/device-tree/model
NVIDIA Jetson AGX Thor Developer Kit

sudo apt search sdkmanager
sdkmanager/now 26.1.0-14392 arm64 [installed,local]
  NVIDIA SDK Manager

scott@chithor:~$ ll /opt/nvidia/sdkmanager/sdkmanager
-rwxr-xr-x 1 root root 120720512 Oct 11 22:57 /opt/nvidia/sdkmanager/sdkmanager

If you have a usb port you could try https://github.com/whitesscott/jetson-live-iso
That needs to be run on an Orin or Thor. I authored it and have used it with included jetson-rescue script to mount & chroot into live Thor Jetpack disk and repair configuration and also to reapply Nvidia kernel packages with sudo apt install --reinstall …

You can install Jetson on the Thor itself, without needing an x86_64 architecture. You could remove the NVMe and connect it to your ARM or RISC-V computer, either with a USB adapter or directly installed on the motherboard. You could then directly burn the JetPack ISO to the NVMe drive using dd.

The main downside to this is that you can’t update the QSPI using this method, so you’d only be able to run a supported JetPack ISO.

Thanks for the suggestions. I should have been clearer about the constraints.

The system contains a single Jetson Thor inside a sealed unit. After deployment there is no physical access: no one can connect a host PC, plug in a USB drive or remove the NVMe. There is no second Jetson to use as a host.

The board also has a small embedded supervisor (Probably an FPGA with ARM or RISC-V processor, not x86) that controls Thor’s power, reset and FORCE_RECOVERY_N, and is connected to Thor’s USB0 recovery port. When the supervisor detects that Thor fails to boot, for example BCM boot is corrupted, it must recover the module autonomously: put Thor in recovery mode and reflash BCM and, if needed, NVMe over USB.

As I understand it, this is currently blocked because:

  • JetPack flashing is not supported from ARM64 hosts, and
  • Applications for flashing are x86-only binary and I didn’t find any documentation for porting it to our processor architectures.

So my questions are:

  1. Is there, or is there planned, an ARM64 (or source-available) build of the related flashing tools for Jetson Thor?
  2. Can NVIDIA provide the RCM protocol specification (under NDA if required) so we can implement a minimal recovery host on our supervisor?
  3. Is there any supported mechanism for a Thor to recover its own BCM without an external host? For example, a golden or recovery image, or an alternative boot source.

Thanks,

Federico

To reiterate. Nvidia has an aarch64 / arm64 sdkmanager

Currently on my Thor Jetpack 7.2.1 where I have sdkmanager installed here’s this:

scott@chithor:~$ sdkmanager --cli

+++++++++++++++++++++++++++++++++++++++++++++++++++
+   Welcome to NVIDIA SDK MANAGER v26.1.0.14392   +
+++++++++++++++++++++++++++++++++++++++++++++++++++

┏ Login ----------------------------------------- ┓
✔ Logging into NVIDIA Developer server...
✔ Loading user information...
  Hello scott :)
✔ Loading server data...
✔ Loading and processing available products...
┗ ----------------------------------------------- ┛

┏ Install options ------------------------------- ┓
- Select action: (Use arrow keys)
❯ Install 
  Download only 
  Uninstall 

and go to this webpage to download arm64:

https://developer.nvidia.com/sdk-manager#getting-started

sdkmanager_26.1.0-14392_arm64.deb
sudo apt install ./sdkmanager_26.1.0-14392_arm64.deb

sdkmanager --cli

It installs to /opt/nvidia/sdkmanager/


This might be useful. On your board confirm Jetpack version and that bootloader is status normal with sudo nvbootctrl dump-slots-info In that state the Thor will boot to the other slot should one be corrupted.

You could check to implement to your environment / use case. Nvidia only enables /opt/nvidia/l4t-bootloader-config/* on IGX boards.

ls /opt/nvidia/l4t-bootloader-config
nv-l4t-bootloader-config.conf  nv-l4t-bootloader-config.service  nv-l4t-bootloader-config.sh  nv-l4t-bootloader-update.sh


# To retrieve and stage  TEGRA_BL_3834_agx.Cap

apt download nvidia-l4t-bootloader

# Or if you are on earlier than 39.2.1 go to https://repo.download.nvidia.com/jetson/
# copy the link to your nvidia-l4t-bootloader_$VERSION_arm64.deb

wget $VERSION-link

dpkg-deb -x nvidia-l4t-bootloader_*_arm64.deb .

sudo mkdir -p /opt/ota_package/t26x
sudo cp -p  ./opt/ota_package/t26x/  /opt/ota_package/t26x/

Thanks for the information.

Our supervisor currently runs Buildroot, not Ubuntu/JetPack, and our ideal solution is a small, self-contained recovery application that doesn’t depend on the JetPack libraries. That’s why understanding the recovery protocol matters to us, especially for a port to RISC-V.

That said, your information about the arm64 SDK Manager, the A/B slot behavior and the bootloader update flow is a good starting point. We’ll try to port it to our system to see what happens.

Thanks again for your help.

We don’t have flash tool for ARM64 or RISC-V host, and no RCM protocol specification to provide.

Thanks