Apple Researchers Publish Minions System: Aggregating Idle Mac GPUs into Virtual Cloud Cuts AI Inference Queue Times from 10 Hours to Minutes

According to a report by tech media outlet AppleInsider on October 15, Apple's research team has published a study on a system called Minions. The proposal aims to aggregate long-idle GPU resources from Mac devices within an enterprise to form a dynamically schedulable "virtual GPU cloud" for running large-scale machine learning inference tasks.

Traditional centralized GPU clusters often face issues such as high costs, slow scaling, and long task queuing times. Minions addresses this bottleneck through four key components: the Minion Server handles task splitting and scheduling; PostgreSQL stores metadata; the Minion Executor, running on macOS, uses local Apple Silicon GPUs to perform inference; and the Minion Watchdog monitors device status and reassigns tasks if devices go offline.

The system employs a "pull-based" model where devices actively claim sub-tasks, using atomic mechanisms to prevent duplicate processing. Apple disclosed that in internal tests, inference tasks that previously queued for over 10 hours due to resource constraints could begin execution within minutes under the Minions architecture, without requiring new hardware purchases.

Currently, this technology remains an internal research and engineering solution rather than a formal consumer product. The research team noted that data confidentiality and governance remain significant concerns when utilizing other employees' devices across departments.