Apple's Minions pools idle Macs into a GPU cloud, with machines recruited over Slack

Apple researchers have described Minions, an internal system that pools the idle graphics processors of Apple Silicon Macs into a shared cloud for machine learning inference. The paper says some jobs that once waited more than ten hours for central GPUs can now start within minutes, and that teams often recruit extra machines simply by asking colleagues on Slack.

A GPU cloud made of office Macs

The paper, posted to arXiv on October 9 and listed on Apple's machine learning research site, starts from a plain observation. Modern Macs carry capable GPUs that sit mostly unused during ordinary work, while centralised GPU clusters are expensive to grow and often oversubscribed.

Minions has four parts: a server that hands out work, a PostgreSQL database that holds all task and machine state, a watchdog, and an executor program that runs on each participating Mac. Inputs and outputs live in Amazon S3. Large jobs are split into subtasks that participating Macs claim from the server.

Two kinds of work are supported. Embedding jobs run open source Hugging Face models locally, and text generation runs through Ollama, the open source tool for running language models on a Mac. Nothing is sent to an outside API.

Built for machines that come and go

The design assumes laptops will vanish without warning. Executors send a heartbeat every five minutes, and the watchdog checks only the machines currently doing work. If one stops answering because of network loss, sleep, shutdown or its owner quitting the program to get the full machine back, its subtask goes back in the queue for another Mac.

Owners also decide where their machine's time goes. Each executor can be set to prefer a particular team, project and task, and falls back to the shared queue when that team has nothing waiting. The paper describes teams temporarily gathering the Macs they manage to push one urgent job ahead.

What the paper does not measure

The results section is short. Apple's table lists queue time falling from hours, up to more than ten, to minutes, new GPU requests going from required to none, and infrastructure cost going from high to zero. It gives no fleet size, throughput, energy figures or comparison against a data center GPU.

The authors also flag a social limit. Some colleagues hesitated to run other teams' data on their machines, so Minions supports separate deployments with their own database and pool of Macs.

The paper ends on a much bigger idea. If idle Macs across organisations, "or potentially across the global community", could be securely federated, the authors write, they could form a large shared GPU pool. For now, the next step they name is adding model training alongside inference.