Apple Builds Virtual GPU Cloud From Idle Macs
Apple cut internal AI queue times by treating hardware it already owns as a pool rather than buying more GPU servers, which is the cheaper path for any company sitting on large fleets of idle machines.
Reporting from 1 source: GIGAZINE.
Apple's research team has developed Minions, a distributed system that gathers idle GPU capacity from Macs across the company and runs AI jobs on them as a virtual GPU cloud. Large jobs are split into small tasks, each handled by a Minion Executor on an Apple Silicon Mac, with results saved to Amazon S3. Apple says waits that once exceeded 10 hours now begin within minutes.
The system splits a large job into small tasks and hands each one to a Mac on the network. A program called Minion Executor pulls a task and runs it on the Apple Silicon GPU, then writes the result to Amazon S3. Each Mac is not carrying part of one model. Each one runs its own task independently.
Because those Macs belong to people, they can sleep, drop off the network, or be reclaimed by their owners. Each machine sends a heartbeat every five minutes, and Minion Watchdog returns unfinished tasks from silent Macs to the queue for another machine to redo. Owners can also set which teams get priority on their GPU.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.