How a Three-Device Home Lab Schedules AI Jobs on a Single-Model GPU

A hobbyist home lab built from a Raspberry Pi 5, a Jetson Orin Nano, and an RTX 4070 Ti Super desktop GPU manages AI workloads on a tight memory budget roughly equivalent to the cost of a mid-range laptop. The Jetson Orin Nano's shared CPU-GPU memory pool can hold only one model at a time, making simultaneous job requests a risk for memory thrashing. To prevent conflicts, all jobs — including chat requests, research tasks, and scheduled scrapes — are routed through a single priority queue with a gateway that admits exactly one job to the Jetson at a time. Priority is respected but non-preemptive, meaning a higher-priority job waits for the current one to finish rather than interrupting it, avoiding wasted work and unstable model states. The RTX machine handles rendering independently but shares the same queue only for a common LLM call where both machines would otherwise compete for the same resource.