Apple Explores Distributing GPU Computing Across Macs

Apple's research team has unveiled a system study called Minions, which seeks to aggregate the idle GPU power of Mac computers into a dynamic virtual GPU cloud. The initiative aims to speed up large-scale machine learning inference tasks by efficiently utilizing existing hardware.

Traditional AI inference often relies on costly, centralized GPU clusters that are slow to scale and can lead to significant queuing times. Many Macs, used for tasks like office work, browsing, and content creation, have GPUs that remain unused for extended periods. Minions aims to address this by harnessing these underutilized computational resources.

The system comprises several components: a Minion Server manages tasks and distributes sub-tasks to devices. A PostgreSQL database stores metadata on task and device status. A Minion Executor on each Mac performs the inference locally using its Apple Silicon GPU. A Minion Watchdog monitors devices for disconnections or sleep states, reallocating unfinished tasks to available machines.

Apple's tests indicate that tasks previously requiring over 10 hours of queue time due to limited centralized GPU resources can now be completed in minutes using Minions. This approach reduces infrastructure costs by eliminating the need for additional GPU hardware. However, the research is currently positioned as an internal corporate solution rather than a consumer product, with considerations for data privacy and governance when using devices across different teams.