With graphics cards in short supply, Apple requisitions its employees’ Macs

Apple’s engineers have just documented Minions, an internal system that puts the company’s idle Mac GPUs to work. The authors summed up the concept in one sentence: “a virtual, dynamic and distributed GPU cloud that pools the idle Apple Silicon GPUs of macOS machines to run large-scale machine learning inference workloads”. Before its deployment, some tasks waited more than ten hours in the queue because no GPUs were available in the centralized pool; they now typically start within minutes.

Four gears and a watchdog

The system relies on four components. A central server, the Minion Server, receives requests from teams and breaks them down into subtasks, whose status is recorded in a PostgreSQL database. On each enrolled Mac, a Minion Executor runs its share of the computation on the local GPU, uploads the result to S3 storage, and reports back to the server.

A watchdog, the Minion Watchdog, monitors the whole system: if a machine stops responding, its tasks are reassigned elsewhere. Dashboards provide real-time visibility into the health of the herd.

Recent Mac mini and Mac Studio models can also pool via RDMA over Thunderbolt 5, combining memory and computing power to load a model too large for a single machine.

Availability over raw power

This approach contrasts with Nvidia GPU clusters, which are optimized for throughput and latency on very large models. Minions primarily targets utilization rates and harnesses idle desktop resources. The system redistributes computing power already available over the internal network, benefiting teams that lack it, rather than targeting the market’s giant models.

Outside Cupertino, services are already allowing individuals to rent the computing power of their Macs to AI companies, sometimes for as much as $200 per month. Apple has just applied the same approach internally, using machines it owns. It’s an ironic situation for the company that designs some of the world’s best inference hardware.

👉🏻 Follow the latest tech news in real time: add 01net to your feeds on Google and subscribe to our WhatsApp channel.

Source: Apple