Apple researchers have unveiled Minions, a system that pools idle GPUs from Apple Silicon macOS devices into a dynamic, distributed virtual GPU cloud for machine learning inference. The paper, published on Apple's machine learning blog, details how the approach addresses the high cost, limited elasticity, and long wait times of centralized GPU infrastructure.
The work comes as demand for GPU-accelerated inference skyrockets, and Apple Silicon Macs—equipped with powerful GPUs—often sit idle during normal use. Minions aims to tap that untapped capacity for large-scale inference workloads without new hardware purchases.
How Minions works as a distributed GPU cloud
Minions aggregates idle GPUs across multiple Apple Silicon macOS devices to execute inference tasks. It provides flexible task execution, fault tolerance, and efficient resource utilization, according to the paper. The system is designed for environments where centralized GPU resources are constrained, allowing teams to leverage existing hardware.
The paper's authors—Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan, and Sean Suchter—note that idle macOS devices can be transformed into a practical distributed GPU inference platform. This could offer a cost-effective alternative to traditional GPU clusters.
Apple's broader AI infrastructure push
Minions is part of Apple's growing machine learning research. The company has also explored running large language models efficiently on Mac using MLX and the neural accelerators in the M5 GPU, aiming to enable private, on-device AI experimentation.
Additionally, Apple has worked on accelerating LLM inference on NVIDIA GPUs with ReDrafter, showing a dual focus on both Apple Silicon and industry-standard hardware. This suggests Apple is thinking broadly about AI infrastructure, from edge devices to cloud-scale inference.
