A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices
Points and comments are a snapshot, not live.
Apple built an internal system that pools idle Mac GPUs for distributed ML inference.
Minions aggregates idle Apple Silicon GPUs across macOS devices on a corporate network to execute machine learning inference workloads. The system uses task-level parallelism, breaking workloads into independent subtasks that run on separate machines, such as generating embeddings for a million documents. It relies on Ollama as a foundation and stores inputs and outputs in S3. The paper is a short write-up of an internal enterprise system already in use at Apple, not a research breakthrough.
What commenters are saying
Commenters note this is not new: similar idle-cycle harvesting has been done for decades with Condor, SETI@Home, and distributed rendering farms. Two camps emerge: those who see it as a pragmatic internal tool for Apple, and those disappointed by the paper's lack of depth and reliance on Ollama. Some highlight practical challenges like tail latencies, fleet heterogeneity, and scheduler management. A few compare it to projects like exo or Salad.