Google is working with the University of California, San Diego (UCSD) to transform old Pixel phones into low-cost computing clusters. According to Google's research, the computing power of 25 to 50 old phones can rival a dual-socket server-grade CPU.The research team first removed unnecessary components on the mobile phone, including the screen, battery, camera, speakers and casing, leaving only the motherboard equipped with the SoC. Then, the Android system was replaced with a Linux distribution common to data centers, and after removing redundant software on the consumer side, orchestration tools such as Kubernetes were deployed.
Benchmark tests show that the single-core SPEC score of a mobile phone three years ago is actually higher than the Asus RS720A-E11 server equipped with dual-channel AMD EPYC and NVIDIA H200.Although the overall computing power of the server is far from that of mobile phones, the leading single-core performance means that old mobile phones are still valuable for specific computing tasks.
In practical applications, a cluster of 20 old mobile phones can support the teaching application of a class of more than 75 people, without relying on the cloud at all.
The research team plans to use 2,000 mobile phones to build a local data center and support the operation of hundreds of classes at the same time. The hardware cost is only a fraction of that of a new server. The advantage is even more obvious in the current context of rising prices of memory and storage chips.
Reusing old equipment is not a new idea. Last year, a research team used four old mobile phones to form a micro data center for underwater monitoring. NASA even installed the 2014 mid-range chip Qualcomm 801 on the Ingenuity Mars helicopter to provide navigation calculations for the Perseverance Mars rover.
For universities and small and medium-sized institutions, old mobile phone clusters provide a feasible way to bypass the sky-high price hardware of technology giants. However, AI hyperscale data centers are unlikely to switch to this solution because reliability and maintenance costs are still shortcomings.
