Apple's Minions: A virtual GPU cloud built from idle Apple Silicon Macs

Minions is a system that pools idle Apple Silicon GPUs on macOS devices into a dynamic, distributed virtual GPU cloud for machine learning inference. It addresses shortcomings of centralized GPU farms - high cost, limited elasticity, and long queue times - by aggregating otherwise-unused GPU capacity across many Macs to run large-scale inference tasks. The system schedules and executes inference workloads across heterogeneous, intermittently available devices, providing flexible task placement, fault tolerance against device churn, and mechanisms to maximize overall GPU utilization without buying additional hardware.

Evaluation and experience with Minions demonstrate that fleets of idle macOS devices can serve as a practical inference platform when centralized GPUs are constrained. By turning personal and office Macs into a pooled resource, teams gain elasticity and cost savings while reducing wait times for inference workloads. The design emphasizes resilient execution across transient nodes, efficient resource packing, and seamless integration with existing inference pipelines, showing that repurposing idle Apple Silicon GPUs is an effective complement to traditional centralized GPU infrastructure.