Apple researchers have introduced Minions, a system designed to aggregate idle Mac GPUs into a dynamic virtual cloud for machine learning inference. This approach aims to utilize existing hardware instead of relying on centralized GPU clusters, which often face high costs and long wait times.
The system uses Minion Executor to run tasks on Apple Silicon GPUs and Minion Watchdog to manage device availability, including handling disconnections or sleep modes. It employs a pull-based model where devices request sub-tasks from a server, ensuring efficient distribution and preventing redundant processing.
Testing showed that tasks previously requiring over 10 hours of queuing on centralized clusters could begin execution within minutes using the Minions framework. The system also includes a Grafana-based dashboard for real-time monitoring of device status and workload distribution.
Currently, Minions is an internal research and engineering project rather than a consumer product. The research team noted that implementing this across departments requires addressing significant challenges related to data privacy and corporate governance.
