Apple introduces Minions system
Apple has published a research paper introducing Minions, a system designed to gather unused GPU power from Macs with Apple Silicon chips. This system allows companies to use the idle processing power of their Macs for tasks like running artificial intelligence models, improving efficiency and reducing the need for dedicated GPU clusters.Minions' four components
The system is built around four main parts. A central server handles incoming requests and breaks them into smaller tasks. Executors on individual Macs process these tasks using the GPU and store the results. A "Watchdog" supervisor monitors the executors and requeues tasks if a Mac becomes unresponsive. A dashboard provides real-time monitoring of the entire system.Limitations of the system
The Minions system allows for the use of GPU power during inference tasks, but it does not support model training at this time. Each Mac must be able to run the full model locally, which limits the types of models that can be used. For example, the recently introduced Muse Spark Open model cannot run on Apple Silicon Macs due to memory constraints.Workloads and performance
The system is designed for specific tasks, such as vector representation models from the Hugging Face Transformers library and text generation via Ollama. However, the research paper does not provide performance metrics like processing speed, energy use, or recovery rates. It also lacks a central scheduler, meaning tasks are not split across multiple machines.Context and broader implications
Apple’s Minions system is part of a broader effort to position Macs as AI-friendly devices. The system aligns with Apple’s recent collaborations, such as its partnership with Meta and the release of the Muse Spark Open model. While the system is not a breakthrough, it could be useful for organizations with many Macs, especially if they have sufficient memory and are willing to share resources securely.
