Repurposing Low-Cost BC-250 Mining APUs for Budget Local LLM Inference

A developer built a budget-friendly ~$300 local AI rig by combining two repurposed AMD BC-250 crypto-mining APUs using llama.cpp's RPC backend and Vulkan graphics driver. The dual-APU setup achieves inference speeds of 60 tokens per second on a 35-billion parameter Qwen model with a 64k context window. This build highlights an exceptionally cost-effective hardware path for local LLM enthusiasts, proving that surplus crypto-mining gear can be repurposed for modern AI workloads. It demonstrates how distributed inference using llama.cpp RPC enables multi-node hardware scaling over standard gigabit Ethernet without requiring high-end enterprise GPUs. The AMD BC-250 boards, which utilize cut-down PlayStation 5 APUs, cost approximately $115 each and offer roughly 27GB of combined VRAM across the two units. Running on Bazzite OS, the system distributes model layers across both APUs connected over a simple 1 Gbps network connection.