Minions: A Dynamic and Distributed Digital GPU Cloud Constructed on Idle Apple Silicon Units

The growing demand for GPU-accelerated machine studying inference has uncovered limitations in centralized GPU infrastructures, together with excessive price, restricted elasticity, and lengthy wait occasions below useful resource constraints. At the identical time, fashionable Apple Silicon–based mostly macOS gadgets are geared up with highly effective GPUs that stay largely idle throughout regular utilization. This paper presents Minions, a dynamic and distributed digital GPU cloud that aggregates idle Apple Silicon GPUs throughout macOS gadgets to execute large-scale machine studying inference workloads. Minions permits groups to leverage present {hardware} with out requiring further GPU procurement. The system supplies versatile GPU inference process execution, fault tolerance, and environment friendly useful resource utilization. Our expertise reveals that idle macOS gadgets may be remodeled right into a sensible distributed GPU inference platform in environments the place centralized GPU assets are constrained.