Apple proposes shared computing system

Apple researchers have introduced a system named Minions, designed to transform a group of Apple Silicon Macs into a shared resource for artificial intelligence tasks. This system enables users to harness the unused graphics processing units (GPUs) of their existing Macs, eliminating the need for extra server infrastructure.

How Minions works

The system relies on a central server where users submit tasks. This server divides the tasks into smaller, independent parts and keeps track of their progress using a PostgreSQL database. Each participating Mac then retrieves a task, downloads the required data, and performs the computation using its GPU.

Supported tasks and data flow

Currently, Minions supports two types of tasks: generating embeddings, which are numerical representations used in search engines to relate documents based on meaning, and running responses from language models using Ollama. These models are downloaded as needed and stored on the Mac for future use. Input data and results are stored temporarily in AWS S3, connecting the local computation to cloud infrastructure.

Distributed approach and similarities

Minions does not combine multiple Macs into a single machine but instead distributes tasks among several computers, each using its own resources. This approach is similar to the Seti@Home program, which also uses distributed computing to analyze data for extraterrestrial signals.

Challenges with personal computers

A potential issue with using personal computers instead of data center servers is that users may shut them down or use them for other purposes. To address this, each Mac regularly sends a "signal of life" to the network. If a machine goes offline or stops the program, the interrupted task is returned to the queue, where another Mac can take over.

Informal organization and task management

The system operates in a somewhat informal way, with colleagues using Slack to request the program be launched and temporarily share their GPU resources. Some tasks that previously took over ten hours on centralized systems can now start in just a few minutes with Minions, though the benefit is more about reducing wait times than increasing speed.

Limitations and context

The researchers did not provide specific numbers of Macs used, detailed performance metrics, or energy consumption data for comparison with traditional GPU servers. The "zero cost" claim refers to not needing to buy new hardware, though electricity, storage, and management still incur costs. Minions is presented as an experience report by Apple researchers rather than a planned macOS feature, emphasizing the use of existing resources before investing in new ones.