THE MACHINE

The lab is a rack in a spare room.

ArkNode runs on one Proxmox host carved into LXC containers — one GPU each, no VMs, no cloud. Every experiment on this site trained, backtested, or failed on the hardware below.

It is deliberately boring: consumer parts, ECC where it counts, a runbook for everything. When the hardware misbehaves, that becomes a post too.

TOPOLOGY
pve01PROXMOX VEEPYC 7443gpu01RTX 4090 · TRAINgpu02RTX 4090 · TRAINgpu03RTX 3090 · INFERgpu04RTX 3090 · EMBEDnas01ZFS · 48 TB
HARDWARE
NODEHARDWAREMEMROLESTATUS
pve01EPYC 7443 · 24C512 GB ECCProxmox VE host● ONLINE
gpu01RTX 409024 GBTraining● ONLINE
gpu02RTX 409024 GBTraining (busy)● BUSY
gpu03RTX 309024 GBInference · evals● ONLINE
gpu04RTX 309024 GBEmbeddings · data● ONLINE
nas01ZFS mirror48 TBStorage● ONLINE

Full runbooks — GPU passthrough, thermals, backup policy — live in the build logs. If it broke at 2 a.m., there’s a post about it.

READ THE BUILD LOGS →