Apple develops Minions system to create virtual GPU cloud from idle Macs

Apple has developed a distributed computing platform called ‘Minions’ to address internal shortages of graphics processing units (GPUs). The system creates a virtual GPU cloud by aggregating the unused Apple Silicon GPU resources from employees' Mac computers across the company.

The architecture consists of a central Minion Server that manages tasks, a PostgreSQL database for metadata, and a software agent called Minion Executor that runs on individual Macs. To ensure reliability, a ‘Minion Watchdog’ monitors the network; if a Mac becomes unavailable due to sleep or disconnection, its tasks are automatically reassigned to other machines. Results are stored in Amazon S3.

Currently, the system supports embedding generation using Hugging Face open-source models and text generation via Ollama. By utilizing existing hardware, Apple has reportedly reduced AI processing wait times from over ten hours to just a few minutes. While the system currently focuses on inference, researchers plan to expand its capabilities to include AI model training in the future.