Google Turns Thousands Of Pixel Phones Into A Low-Carbon Data Center

hero oldphones
Nothing like some positive news to start out the week: In search for a solution to counter the massive carbon footprint of data centers, researchers at the University of California San Diego, with support from Google, are giving retired smartphones a second life as general-purpose cloud computing servers, rather than buying new expensive hardware.

On average, consumers swap out their smartphones every four years. The thing is, many of these discarded devices still possess core internal components that remain robust. In fact, benchmark testing reveals a surprising fact: processors inside smartphones from just three years ago frequently deliver single-core performance that equals or beats traditional, high-end multi-core data center servers. Although a smartphone lacks the vast memory capacity and dozens of simultaneous multi-threaded cores found in a standard server rack, its raw per-core computing power can be incredibly potent.

To tap into this dormant potential, UCSD researchers stripped the retired phones of components that are either hazardous or useless in a data center context, e.g. the glass displays, cameras, chassis, and lithium-ion batteries. By leaving (and reusing) the motherboard and SoC alone, one is able to mitigate the environmental impact of producing brand new chips, as manufacturing these components accounts for roughly half of a smartphone's entire lifetime embodied carbon footprint. 

Next, Android OS is replaced with a general-purpose Linux distro. This software update disables mobile-specific limitations, such as background application throttling, allowing the hardware to run at max performance. Because it takes between 25 and 50 smartphones to match the total compute throughput of a single dual-socket server, the team utilizes Kubernetes orchestration software to bind the individual motherboards into cohesive, self-managing clusters.

phone cluster latency1
Cluster latency and throughput for a 20 phone cluster running a parallel grading task. (Credit: Google/UCSD)

The practical applications of this platform are already being demonstrated on campus. Many routine academic tasks, like hosting digital notebooks or running automated grading software for computer science classes, require minimal memory and easily fit within a single smartphone's capacity. Early testing with a modest 20-phone cluster successfully handled peak assignment submissions for a class of over 75 students, achieving lower latency than standard commercial cloud backends.

Building on this success, the university is preparing to launch a massive, 2,000-Pixel phone computing cluster. Expected to go live ahead of the upcoming academic year, this makeshift data center will be capable of supporting a hundred computer science courses simultaneously. 

And sure, tech companies chasing massive AI models are unlikely to swap out specialized hardware for salvaged phone parts, but this recycling model proves that an incredibly cheap, highly sustainable alternative for educational institutions and smaller organizations is totally viable, reliable, and can help us sleep better at night.

Main image credit: Jan Antonin Kolar via Unsplash
AL

Aaron Leong

Tech enthusiast, YouTuber, engineer, rock climber, family guy. 'Nuff said.