Apple has detailed a new system, called Minions, that uses the idle GPUs found in Apple Silicon Macs to create a virtual GPU cloud that combines untapped processing power to chew through complex AI workloads.
Described by Apple as a dynamic virtual GPU cloud that "aggregates idle Apple Silicon GPUs" across Macs, Minions can aid teams that need to work with large-scale machine learning inference workloads. Apple's researchers have been using Minions in-house already.
The theory is a simple one. Macs have powerful GPUs that spend most of their time idle, especially during everyday computing tasks.
Minions is designed to leverage the GPU processing power that isn't being used, turning otherwise underused Macs into a cluster of AI-ready machines.
Both the Mac mini and Mac Studio support clustering via the Remote Direct Memory Access (RDMA) system using Thunderbolt 5 connections. These high-bandwidth cables allow compute and memory to be combined from multiple machines, putting them to work on a single workload.
Minions is an expansion of that concept.
Components of a Minion
Apple's documentation explains that there are four key components required for Minions to function.
First, a Minions Server acts as the central coordinator, handling all of the requests via a web-based interface and APIs. Once a job has been handed to the Minions Server, everything is handed off and queued into a PostgreSQL metadata store.
Minion Executors run on the macOS devices whose GPU cycles are to be harvested for the Minions system. These Executors execute subtasks locally using the Apple Silicon GPUs and then write their results to storage.
If a failure occurs, such as during a connectivity failure or power loss, the Minion Watchdog re-queues the failed job for execution on another machine.
The whole process is fed back into a dashboard for real-time monitoring.
Apple argues that, in its experience, Minion has proven that macOS devices can be turned into a distributed GPU inference platform when needed.
Apple also says that, in its testing, some inference jobs that previously took more than 10 hours have been slashed to mere minutes. And all without the need for additional infrastructure and with an easy path to increased scale.
When that might be needed will differ from one use case to the next. But it's easy to imagine an office full of Macs having their GPU cycles harvested in this way, as the "Minions" method does not appear to require Thunderbolt.
For businesses, the benefits of something like Minions are obvious. Slashing processing time without having to buy additional hardware is always a plus, especially as prices continue to skyrocket.
Some businesses may also see this as a good reason to buy Apple devices come refresh time. Something Apple will surely encourage.
