Google discussed the development of a low-carbon computing platform at the University of California at San Diego, utilizing two thousand motherboards from used Google Pixel smartphones. By repurposing used smartphones instead of purchasing new equipment, researchers and students can create a functioning cluster for experiments at minimal cost. This initiative is seen as a way to decrease carbon emissions by eliminating the need to produce new servers for computing infrastructure.
Despite the common practice of users upgrading their smartphones every four years for social reasons rather than necessity, many devices with powerful processors are left unused. Google suggests giving these devices a second life by constructing clusters for cloud computing.
The performance of processor cores in modern smartphones when handling single-threaded tasks often matches or surpasses that of modern server cores. For example, in various single-threaded tests from SPEC 2017, the high-performance cores of the 2023 Pixel Fold smartphone outperformed the cores of a common server from ASUS. Though server processors have numerous high-performance, multi-threaded cores, smartphone SoCs combine a smaller number of high-performance and power-efficient cores. Taking core count into consideration, the computing power of 25-50 smartphones is equated to that of a modern server.
The cluster was assembled solely using motherboards from smartphones, with a Linux server distribution replacing the Android system. Due to the limited RAM on smartphones, the cluster can only handle tasks that require the memory available on one device. To streamline task orchestration, the smartphones forming the cluster are divided into groups of 25-50 devices, each running applications in isolated containers managed through Kubernetes.
The cluster is expected to be operational by the fall, supporting courses on parallel computing and system programming. Tests have shown that a cluster of 20 smartphones can manage the workload of courses for over 75 students, with lower latencies than a previous AWS-based backend. Once completed, the cluster of 2,000 smartphones should offer computing resources for hundreds of courses.