Apple's Minions Turns Idle Macs Into a GPU Cloud, Cutting 10-Hour Queues
Apple researchers describe an internal system that pools idle Mac GPUs across the office network, cutting inference queues from 10+ hours to minutes.
Takeaways
- Apple published Minions, pooling idle Mac GPUs into a distributed inference cloud.
- Internal queues dropped from 10+ hours to minutes on constrained GPU workloads.
- Four components: Minion Server, PostgreSQL store, Executor on each Mac, and Watchdog for fault tolerance.
- Executors pull subtasks atomically, run inference locally (reportedly via Ollama), and write results back.
- Inference-only today; authors suggest the scheme could eventually scale globally and support training.
- No code, no throughput benchmarks, and no comparison to Nvidia clusters are published.
Apple researchers have published a three-page paper describing Minions, an internal system that distributes machine-learning inference across idle Apple Silicon devices. The authors report that jobs previously blocked for more than 10 hours in company GPU queues can begin within minutes on the Mac fleet.
Minions targets a common capacity problem: centralized accelerators are expensive, heavily shared, and slow to expand, while employee laptops and desktops leave much of their GPU capacity unused. Apple has presented the system as a research design, with no public code release or product announcement.
Why inference jobs get stuck
Machine-learning inference runs a trained model to generate predictions, embeddings, rankings, or evaluations. Large organizations often send both inference and training jobs to the same centralized GPU clusters, where batch work can wait behind longer or higher-priority tasks.
Apple Silicon provides a second pool of accelerators already deployed across employee Macs. Its unified memory architecture lets the CPU and GPU access the same memory pool, which supports models that fit within each device’s available capacity. Interactive development rarely keeps the GPU fully occupied, leaving periods that Minions can use for background work.
Pull scheduling for transient workers
Minions uses a pull-based work queue suited to machines whose availability changes throughout the day. Each Mac requests work while its executor is active, PostgreSQL serializes task claims, and a watchdog recovers work abandoned after a laptop sleeps, disconnects, or reboots.