Apple builds Minions, a distributed GPU cloud from idle Macs for AI inference

Apple researchers detailed Minions, a system that pools idle Apple Silicon GPUs across macOS devices into a virtual GPU cloud for large-scale ML inference. Inference jobs that once queued for over ten hours now start within minutes.

• Minions aggregates idle Apple Silicon GPUs into a shared inference pool
• Inference queue times dropped from over 10 hours to minutes
• The system uses Minion Server, PostgreSQL, Executor and Watchdog
• Mac mini and Mac Studio support RDMA clustering over Thunderbolt 5