Apple's Minions Research Aggregates Idle Mac GPUs for Distributed AI Tasks
Apple's research team unveiled Minions, a system that aggregates idle Mac GPUs into a dynamic virtual cloud to run large-scale machine learning inference tasks. The architecture utilizes a pull-based task model with components like Minion Server, PostgreSQL, Minion Executor, and Minion Watchdog to coordinate and execute workloads. This research offers a novel approach to significantly reduce infrastructure costs and bypass centralized GPU resource bottlenecks without purchasing additional hardware. It effectively leverages existing enterprise hardware assets that are otherwise sitting idle during routine office and creative scenarios. In tests, Minions drastically cut down wait times for inference tasks from over 10 hours down to minutes by distributing work across macOS devices. However, the system is currently designed as an internal research and engineering solution, with data privacy and cross-departmental governance remaining key considerations.
## BACKGROUND
Distributed computing connects multiple computer servers or devices across a network to enable large workloads to take advantage of all available computational resources. AI inference workloads represent the execution phase of artificial intelligence models, where trained models process real-time data inputs to make predictions or decisions.