How cool are the Nvidia Jetson series of “Advanced AI Embedded Systems”! They pack all kinds of potential for machine vision applications, the features of which I won’t go into here. Mainly because I cannot even understand most of them. But hey, AI sounds pretty cool. For me, Tensor Cores conjure up images of Tony Starks nuclear power source. While CUDA cores sound really cute. ARM must be because they cost an arm and a leg, amirite?
But do you know what does not sound cool and cute? Releasing a software update that renders the devices GPIO that is available on the units 40 pin expansion header useless. Who uses GPIO anyway. It is not like this computer is marketed as an embedded system or anything. Why test it before a software release? It is not like Nvidia is a large company with massive resources retailing a small computer system for more than $2000AUD which if you ask me, IS the price of an arm AND A LEG!
Yes I am a little bitter…
Anyway I am writing this because I lost three days of my life trying to get the GPIO on my Jetson AGX Xavier Developer Kit working only to find out from a forum post that “There is a known issue in jp5.0 GPIO” and that it is not currently possible to use Linux sysfs and Jetson.GPIO. Supposedly you can get it working by manipulating the GPIO register addresses directly, but good luck with that. The Xavier SoC Technical Reference Manual is horrendous, and as far as I can tell it is missing information that would allow the development of a bare metal Hardware Abstraction Layer using mmap (some people have had success doing this on the Nvidia Jetson Nano). Hopefully by the time you are reading this the next release of JetPack is out and this issue no longer exists. But currently the issue does exist.
Here are some tips I have compiled that might save you some time getting the thing up and running:
- Do not use JetPack 5 or 5.0.1. Give JetPack 4.6.1 a go. GPIO functionality does not yet work with JetPack 5, and most likely other peripherals do not work as well. I sense that there is a lot wrong with JetPack 5 on the Jetson AGX Xavier and it should simply not be used yet.
- You can only setup and flash the Jetson AGX Xavier with an Ubuntu/Red Hat/CentOS operating system as that is all that is supported by Nvidia SDK Manager. Plan accordingly.
- If you try to use a virtual machine (like me) to run SDK Manager, use VMWare. Virtual Box seems to not work when flashing the Jetsons. Possibly due to the way the USB device dropouts are handled by Virtual Box. I didn’t give Docker a go, it is probably the better option.
- Make sure you use an Ubuntu 18 LTS operating system to install SDK Manager flash the Jetson AGX Xavier because because a newer Ubuntu LTS cannot install older JetPack versions on the Jetson AGX Xavier for some reason (SDK Manager greys out the options). Disregard this if you plan on using JetPack 5.
- Make sure your Ubuntu operating system has around 70GB of disk space (or it can be expanded in the case of a virtual machine). SDK Manager needs a lot of storage space to do it’s thing. I had to recreate my virtual machine twice due to this oversight.
- It can be a pain sometimes to detect the Xavier with Nvidia SDK manager when using VMWare. I had success by plugging into the USBC port on the Jetson and using USB2.0 compatibility in VMWare.
Hi, its really helpful. Bubi wanna ask, for the host device, is is a must to have cuda gpu? I am falshing jetson xavier with vmware as a host (using amd radeon gpu).
Thank you again
Hi!
I am not perfectly sure what you mean, but you can flash the Jetson using an Ubuntu VM that has any hardware, processor, whatever.
Hope this helps.