Apple Turned Idle Macs into 'Minions' to Work on AI
This approach allows the company to solve specific inference tasks more quickly
The Apple Minions system combines available graphics processors from Mac computers and distributes AI workloads among them. Its goal is to more efficiently utilize unused resources within the company, rather than just running large models.
Mac mini and Mac Studio support clustering via RDMA over Thunderbolt 5. This enables the aggregation of computational power and memory from multiple devices, including for models that do not fit on a single computer.
The platform consists of Minion Server, PostgreSQL Metadata Store, Minion Executor, and Minion Watchdog. The server receives requests and splits them into subtasks, the Executor runs inference and saves data to S3, while the Watchdog reassigns tasks to other Macs in case of failures. Work monitoring is conducted through Grafana dashboards. Unlike Nvidia, Apple focuses on the availability of free resources rather than on bandwidth and low latency when working with large models.
