I started working with my Orin NX 16 GB. I found the need for an dedicated x86 or arm64 host unconvinient. So I created a PXE bootable image, that works at least for Orin NX 16 GB, from which I can install without the L4T JetPack toolchain or NVidias SDK Manager.
You can find it here https://github.com/j0ju/sbc-fw-alchemy along side with my TuringPi2 alpine image. The images to build are:
* output/l4t-jetpack7.2.1.initramfs.xz
* output/l4t-jetpack7.2.1.rootfs.tar.zst
* output/l4t-jetpack7.2.1.vmlinuz
Instead of NVIDIAs initramfs I use for the installed system the ones generated by initramfs-tools.
For TP2 the serial console on /dev/ttyTHS1 is enabled, unfortunatly you cannot put the UEFI & Kernel console to anything else than /dev/ttyTCU0.
More specific docs for L4T on TP2 see:
https://github.com/j0ju/sbc-fw-alchemy/blob/main/docs/TuringPi2-Alpine/L4T.Install.md and https://github.com/j0ju/sbc-fw-alchemy/blob/main/docs/L4T/braindump.md
j0ju
08/30/2026, 5:32 PM
This is for now only tested on Orin NX 16GB, any patches are welcome for other tegra modules.
I tested 3d acceleration and Ollama it worked with more than 0.1 TOPS 😉
I need to do some more test and benchmarking and to see if I missed anything else.
For now I changed to this OrinNX with 16GB in Slot1 as Desktop. Box64 runs.
Will do some Pytorch tests in the next weeks depending on the time.
j0ju
08/30/2026, 7:47 PM
The Nvidia Orin is also detecting a Bfluefield card connected to the NVMe 4x connector at the bottom (https://discord.com/channels/754950670175436841/1030129990211219486/1496511923808632933) in contrast to RK1 or CM4/5. This allows more flexible expeirments like a 5 or even 6 node arm cluster cluster, extended network connectivity/bandwidth or the use of Snap to attach more storage via the Bluefield as PCIe virtual function (NVMe or virtio)
l
Leander
08/31/2026, 4:27 PM
Nice! That is quite impressive! Little bit overpowered but it looks like a nice way to chain Turing pi 2 clusters