Retired Phone Forest: Google Plans to Use 2,000 Old Pixels to Power 50 Server Nodes

According to Dynamicz Beating monitoring, Google researchers are collaborating with a team from the University of California, San Diego (UCSD) to explore phone cluster computing. They plan to reassemble the motherboards of 2000 retired Pixel smartphones into a low-carbon cloud data center, providing students and teachers with cost-effective, eco-friendly cloud computing resources.

The typical smartphone refresh cycle is around four years, but the computing cores of retired devices remain intact. The CPU core performance is comparable to modern server cores and even demonstrates equivalent performance in some single-thread tests of SPEC CPU 2017. To adapt the smartphones for the data center environment, the research team stripped away non-essential components such as screens, batteries, casings, and cameras, retaining only the motherboard. This approach is taken because the motherboard manufacturing process accounts for a significant portion of the device's embodied carbon emissions and eliminates potential safety risks associated with batteries in data centers.

On the software side, the team replaced the original Android user space with a generic Linux distribution to bypass memory constraints and protection mechanisms like the "Low Memory Killer" designed for mobile devices. For hardware management, clusters of 25 to 50 phone motherboards are autonomously managed and containerized applications are scheduled using Kubernetes. Based on SPEC CPU 2017 test results, clusters composed of 25 to 50 phone motherboards are nearing the CPU throughput of a traditional server.

Preliminary experiments indicate that a microcluster of 20 phones can smoothly handle peak homework submission times for over 75 students in a parallel computing course. The evaluation tasks take about 50 seconds to complete, with homework grading latency even lower than AWS cloud instances. The team estimates that the overall computing power of a system made up of 2000 phones is equivalent to 50 traditional servers, enabling support for over a hundred system programming and parallel computing courses for teaching and research. The entire system is expected to go live in the fall of 2026 as a long-term testing platform to assess the hardware reliability of consumer-grade hardware under sustained heavy workloads.