iOS as Acceleration

As machine learning models continue to grow in size and power, limited access to sufficient compute resources continues to pose a significant barrier to the utilization of more advanced machine learning by those with weaker setups. We explore the potential for common iOS devices to act as supplementary resources in less powerful compute setups to accelerate model training and inference. iOS devices are equipped with surprisingly powerful processors but face limitations such as memory constraints, thermal throttling, and OS sandboxing. In experimentation, though our iOS device was not power enough to contribute directly to more significant transformer training or inference, via distributed pipelining and parallelism we achieve practical benefits in accelerating certain agentic LRM tool-usage and smaller model training and inference. We demonstrate a theoretical near-elimination of tool-usage-time overhead when agentic LRMs can continue reasoning without needing immediate tool results, parallelizing computation onto an iOS worker. We also demonstrate a 32-36\% base decrease in time per batch for training and inference of the smaller 21.8m ResNet-34 by using an iOS worker as a distributed pipeline stage. These findings highlight the potential for commonplace handheld devices to provide meaningful acceleration in compute workloads.