Apple has built a distributed system called Minions that turns idle GPUs in Macs into a virtual GPU cloud for large machine-learning inference workloads. The company's researchers describe it as a dynamic pool that aggregates Apple Silicon GPUs across macOS devices, making compute resources more readily available to Apple's own teams. Before adopting Minions, some inference jobs reportedly faced queue times exceeding ten hours due to limited centralized GPU availability; with the system, tasks can often begin execution within minutes.
The system relies on support in Apple's recent Mac mini and Mac Studio for clustering via Remote Direct Memory Access over Thunderbolt 5, allowing compute and memory across devices to be combined. Minions consists of four main components: a Minion Server for coordination, a PostgreSQL metadata store, a Minion Executor that runs inference tasks on each Mac's GPU, and a Minion Watchdog that monitors for failures and reassigns tasks as needed. Results are uploaded to S3, and database statistics feed into Grafana dashboards for real-time visibility.
Apple's approach is primarily about redistributing idle resources across a network rather than pushing maximum throughput for very large models. As the researchers note, it is not the same as an Nvidia cluster, which focuses on latency, throughput, and efficient GPU utilization for large models. Minions, by contrast, aims to improve hardware availability and utilization for inference work across existing Mac hardware.