Apple's Minions system pools idle Mac GPUs for distributed AI inference
Apple researchers developed Minions, a distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices. The system is intended to reduce long queue times for machine-learning inference by letting jobs start on available Macs within minutes. It uses components including a server, metadata store, executor, and watchdog, and can combine resources from clustered Mac mini and Mac Studio machines.
Apple's researchers built Minions to make better use of spare graphics processors on Macs running macOS. The setup creates a cloud-like pool of virtual GPUs that can draw on Apple Silicon chips when they would otherwise sit unused, helping inference work begin quickly instead of waiting in long centralized queues.
The architecture includes a coordinating server, a PostgreSQL metadata store, an executor, and a watchdog. The executor performs inference using a Mac's GPU, sends outputs to S3, and reports status; the server tracks tasks and metadata. The watchdog detects failures, such as an unresponsive Mac, and reassigns work. Newer Mac mini and Mac Studio machines can also be clustered through RDMA over Thunderbolt 5.
If similar pooling spreads, smaller teams and researchers without dedicated GPU clusters could gain faster access to inference capacity, potentially shortening experimentation cycles. Employees sharing Mac fleets may see idle hardware become more productive, though scheduling, privacy, and reliability concerns could arise when machines join shared pools. The approach may also influence how organizations