Overview
- On Monday researchers from UC San Diego, working with Google Research, reported that a 20‑phone cluster has already supported an application used by more than 75 students and validated strong per‑core benchmark results against some server cores.
- The team converts retired Pixels by removing batteries, screens and peripherals, loading a general‑purpose Linux distribution and using Kubernetes to treat each phone as an independent compute node.
- Researchers say about 25 to 50 phones can match the aggregate compute of a single dual‑socket server CPU and they plan to scale the system to roughly 2,000 phones expected in fall 2026 to test durability and sustained operation.
- The approach trades standardized server reliability and peak performance for lower cost and lower embodied carbon, and it is not meant to replace hyperscale GPU farms used for frontier AI training.
- Prior prototypes showed very low assembly costs and practical uses for teaching, research and edge deployments, so the project could cut electronic waste and give universities affordable on‑premise compute while raising new maintenance and orchestration challenges.