The Pulse
Apple’s Minions Turns Idle Macs Into Shared GPU Capacity
Apple researchers describe Minions, a system that pools idle Apple Silicon Macs to run machine-learning inference tasks. The paper reports that some jobs previously queued for more than 10 hours can begin within minutes.
AI.info Team · Written with AI ·
Long GPU queues meet idle Macs
Apple researchers present Minions as an answer to a specific computing mismatch: some inference jobs wait hours for centralized GPUs, while Apple Silicon Macs often have GPU capacity sitting unused. The system gathers those machines into a shared pool for machine-learning inference, using hardware teams already have rather than requiring new GPU purchases. The researchers say some jobs that previously waited more than 10 hours can now begin within minutes.
Minions is not a public cloud service or a claim that a collection of Macs can replace a dedicated GPU cluster. The paper describes a system used across multiple teams, in settings where central GPU resources are constrained. Its results are operational observations, not a benchmark comparing Mac-based inference speed or output quality with commercial GPU services.
Executors pull work—and can disappear
The system divides submitted jobs into smaller tasks and records their status in PostgreSQL. Minion Executors running on Apple Silicon Macs request eligible work from a central server, run it locally, and send results to an object store such as Amazon S3. For supported workloads, the paper names embedding generation with open-source Hugging Face transformer models and language-model completion through Ollama.
That design accommodates machines that may go offline without warning. Executors send heartbeats every five minutes; if a device sleeps, loses its network connection, or is shut down, a watchdog releases its unfinished task so another executor can take it. The approach avoids treating every Mac as continuously available, but it also means task completion depends on enough participating machines being online and able to take work.
Teams keep control over participation
Executors can be configured to favor a particular team, project, or task, while falling back to the general queue when no preferred work is available. The paper says device owners can set those preferences, and describes teams temporarily coordinating participation through Slack messages. That makes the pool flexible, but also points to a practical limit: adding capacity may depend on colleagues choosing to run the executor.
Confidentiality is another constraint the researchers identify. Organizations may hesitate to run workloads on devices outside their immediate control, so Minions supports team- or organization-scoped deployments with separate PostgreSQL databases and isolated executor pools. The paper describes those controls, but does not provide a detailed security evaluation or quantify how much data exposure they prevent.
Minutes instead of hours, in one reported deployment
In its operational summary, the paper lists GPU queues of up to 10 or more hours before Minions and minutes with the system. It also records no new GPU requests, zero infrastructure cost, and “organic” rather than manual scaling. Those figures describe the teams’ reported experience; the paper gives no device count, aggregate throughput, dollar savings, or controlled comparison that would show how the results vary across deployments.
The authors frame Minions as a supplement for organizations with underused Apple Silicon devices and constrained centralized capacity, not as a universal replacement for conventional GPU infrastructure. Their next stated step is support for distributed model training, beyond the inference workloads the current system handles. For now, the paper’s measured claim is narrower: some jobs that waited over 10 hours for centralized GPUs began within minutes when idle Macs joined the pool.
“Minions consists of four primary components: Minion Server, PostgreSQL Metadata Store, Minion Executor, and Minion Watchdog.”
The paper’s authors, Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter, write.