Apple has developed a system that makes unused GPU capacity on Macs available for AI applications. Apple calls the technology Minions. The system distributes computing tasks across multiple devices equipped with Apple’s own chips. With Minions, companies can use their hardware more effectively and rely less on dedicated GPU servers.
In an article, six Apple researchers explain how they pooled the computing power of various Macs into a virtual cloud. The project was based on the fact that GPUs in many Macs remain largely underutilized during normal use. Users could leverage this spare capacity to run AI workloads.
Minions identifies available computing power and automatically distributes tasks across the devices. The system takes into account the fact that some Macs may be temporarily unavailable or are currently operating at maximum capacity.
AI inference without additional GPU servers
With Minions, Apple is primarily focusing on inference. Unlike training LLMs, many inference tasks can be processed independently. This makes them suitable for a distributed infrastructure. Apple uses Ollama for the local execution of language models. The company also uses PostgreSQL to coordinate tasks and Amazon S3 for storage.
The researchers note that several internal Apple teams are already using Minions. In practice, wait times for GPU resources have reportedly dropped from more than ten hours to a few minutes.
The article makes it clear that Minions does not turn multiple Macs into a single powerful GPU with shared memory. Instead, the software distributes independent computational tasks across different devices.
According to the researchers, data security is a challenge; after all, when computations run on devices from other teams, sensitive information must be protected.
Apple is still investigating whether the technology can train AI models. In the long term, devices from different organizations might be able to pool their unused GPU resources.