The command you’re looking for is btrfs send. See man btrfs-send.

I know of at least one tool, btrbk, which automates both automatic periodic snapshots and incremental sync, but here’s an example manual process so you can know the basic idea. Run all this in a root shell or sudo.

As initial setup:

• Create a btrfs filesystem on the sender drive and another on the receiver drive. No need to link them or sync anything yet, although the receiver’s filesystem does need to be large enough to actually accept your syncs.
• Use btrfs subvolume create /mnt/mybtrfs/stuff on the sender, substituting the actual mount point of your btrfs filesystem and the name you want to use for a subvolume under it.
• Put all the data you care about inside that subvolume. You can mount the filesystem with a mount option like -o subvol=stuff if you want to treat the subvolume as its own separate mount from its parent.
• Make a snapshot of that subvolume. Name it whatever you want, but something simple and consistent is probably best. Something like mkdir /mnt/mybtrfs/snapshots; btrfs subvolume snapshot /mnt/mybtrfs/stuff /mnt/mybtrfs/snapshots/stuff-20250511.
• If the receiver is a separate computer, make sure it’s booted up and running an SSH server. If you’re sending to another drive on the same system, make sure it’s connected and mounted.
• Send/copy the entire contents of the snapshot with a command like btrfs send /mnt/mybtrfs/snapshots/stuff-20250511 | btrfs receive /mnt/backup. You can run btrfs receive through SSH if the receiver is a separate system.

For incremental syncs after that:

• Make another separate snapshot and make sure not to delete or erase the previous one: btrfs subvolume snapshot /mnt/mybtrfs/stuff /mnt/mybtrfs/snapshots/stuff-20250518.
• Use another send command, this time using the -p option to specify a subvolume of the last successful sync to make it incremental. btrfs send -p /mnt/mybtrfs/snapshots/stuff-20250511 /mnt/mybtrfs/snapshots/stuff-20250518 | btrfs receive /mnt/backup.

If you want to script a process like this, make sure the receiver stores the name of the latest synced snapshot somewhere only after the receive completes successfully, so that you aren’t trying to do incremental syncs based on a parent that didn’t finish syncing.

This seems fairly similar in capability to the Anbernic RG351P, which uses the same RK3326 processor (and by extension the same GPU), similarly has 1GB RAM, similarly has two sticks, and similarly has a conspicuous lack of integrated Wi-Fi. Biggest difference is the screen and body: 351 has a 3.5" 480x320 screen and this is 4" 800x680.

A bunch of Powkiddy devices use this processor too, like the Powkiddy RGB10, RGB20 and V10. Reviews for any of them should tell you how well this’ll perform.

“Dynamically compiled” and dynamic linking are very different things, and in turn dynamic linking is completely different from system calls and inter-process communication. I’m no emulation expert but I’m pretty sure you can’t just swap out a dynamically linked library for a different architecture’s build for it at link time and expect the ABI to somehow work out, unless you only do this with a small few manually vetted libraries where you can clean up the ABI. Calling into drivers or communicating with other processes that run as the native architecture is generally fine, at least.

I don’t know how much Asahi makes use of the capability (if at all), but Apple’s M series processors add special architecture extensions that makes x86 emulation be able to perform much better than on any other ARM system.

I wouldn’t deny that you can get a lot of things playable enough, but this is very much not hardware you get for the purpose of gaming: getting a CPU and motherboard combo that costs $1440 (64-core 2.2GHz) or $2350 (128-core 2.6GHz) that performs substantially worse at most games than a $300 Ryzen CPU+motherboard combo (and has GPU compatibility quirks to boot) will be very disappointing if that’s what you want it for. Though the same could to a lesser extent be said even about x86 workstations that prioritize core count like Xeon/Epyc/Threadripper. For compiling code, running automated tests, and other highly threaded workloads, this hardware is quite a treat.

With one of these Altra CPUs (Q64-22), I can compile the Linux kernel (defconfig aarch64 with modules on GCC 15.1) in 3m8s with -j64. Really great for compiling, and much lower power draw than any x86 system with a comparable core count. Idles at 68W full system power, pulls 130W when all cores are under full load. Pulling out some of my 4 RAM sticks can drive that down a lot more than you’d expect for just RAM. lm_sensors claims the “CPU Power” is 16W and 56W in those two situations.

Should be awful for gaming. It’s possible to run x86 things with emulation, sure, but performance (especially single-thread) suffers a lot. I run a few containers where the performance hit really doesn’t matter through qemu.

Ampere has a weird PCIe bug that results in either outright incompatibility or a video output filled with strange artifacts/distortion for the vast majority of GPUs, with the known good selection that aren’t bugged being only a few select Nvidia ones. I don’t happen to have any of those Nvidia cards but this workstation includes one. Other non-GPU PCIe things like NICs, NVMe, and SAS storage controllers work great, with tons of PCIe lanes.

Depends on what you consider self-hosted. Web applications I use over LAN include Home Assistant, NextRSS, Syncthing, cockpit-machines (VM host), and media stuff (Jellyfin, Kavita, etc). Without web UI, I also run servers for NFS, SMB, and Joplin sync. Nothing but a Wireguard VPN is public-facing; I generally only use it for SSH and file transfer but can access anything else through it.

I’ve had NextCloud running for a year or two but honestly don’t see much point and will probably uninstall it.

I’ve been planning to someday also try out Immich (photo sync), Radicale (calendar), ntfy.sh, paperless-ngx, ArchiveBox (web archive), Tube Archivist (YouTube archive), and Frigate NVR.