This stuff never ceases to amaze me - Haruki Nakamara's mechanical paper animals: https://geecr.com/product/nakamura-paper-toys
Though ITX 3600 + 3060 did sound like a cool idea...
German language is magical: https://youtu.be/gG62zay3kck (feat. Rhabarberbarbara)
https://youtu.be/eZFWDh8ZjFw don't judge me, I love it 😄
Apparently the `--stress` parameter to https://github.com/ComputationalRadiationPhysics/cuda_memtest is what I was missing to diagnose the failing GPU, it identified a memory issue pretty quickly. Now to figure out if there is a Linux BadRAM alternative for GPUs...
Soooo, I've knowingly bought a cheap second-hand 1080Ti that's a bit "broken" - visual artefacts pop up on the screen occasionally, and I managed to lock up the system too (Doom on 1440p ultra - ran smoothly until it didn't). I ran cuda_memcheck and it showed no issues. Are there any GPU diagnostic tools that could identify what's wrong?
What ended up working is a) Setting "Table = off" in the Wireguard config, not to create the "move all traffic through Wireguard" routing rules, b) manually route traffic coming from wg0: "ip route add default via $WG0_IP dev wg0 table 1000; ip rule add from $WG0_IP table 1000"
The problem I had is that I need the "AllowedIPs = 0.0.0.0/0" since I want to be able to serve traffic "from everywhere" (federation FTW!). Unfortunately, when using 'wg-quick' it means that it will set up special routing rules to redirect *all* (even outgoing) traffic through Wireguard, and thus VPS. I noticed when Steam downloads became slow :) (yes, this server is also a gaming machine).