ព័ត៌មានសំខាន់
Apple researchers have introduced a system called Minions, which transforms idle Macs into a virtual GPU cloud to manage internal AI workloads. By distributing tasks across a network of machines, the system significantly reduces wait times for AI jobs that previously faced queues lasting over 10 hours, often completing them in mere minutes.
The architecture utilizes a pull-based execution model consisting of a Minion Server, Executor, and Watchdog. Tasks are split into subtasks stored in a PostgreSQL database, which individual Macs then retrieve via REST APIs. Each Mac processes its assigned subtask locally using Apple Silicon for AI inference, specifically handling embeddings or text generation via Ollama.
Reliability is managed by a Watchdog that monitors heartbeats every five minutes. If a machine goes offline or a user reclaims their device, unfinished subtasks are returned to the queue. While the system offers zero infrastructure costs and organic scaling, researchers note that it does not account for hardware wear or electricity consumption.
Currently limited to internal use, Minions addresses Apple's concerns regarding constrained computing resources and supply chain bottlenecks. Future research aims to extend this framework to distributed model training, with the potential for securely federating idle hardware across global communities to create a shared GPU pool.
