Junkyard Computing: Repurposing Discarded Smartphones to Minimize Carbon

Manufacturing accounts for 70–85% of a smartphone's lifetime carbon, and 150 million phones are discarded yearly in the US alone — so this paper builds a cloudlet of $65 Pixel 3As, defines carbon-per-unit-of-work to replace energy metrics, and measures ten old phones serving real microservices at 9.8–18.9× better carbon efficiency than the equivalent AWS instance.

Why this problem is worth solving

The site’s Carbon Explorer and Ecovisor explainers cover the operational side of computing’s carbon problem — when and where to draw clean energy. This paper attacks the other, often larger half:

The scale is what makes this a systems problem rather than a curiosity. 150 million phones are discarded per year in the United States alone; most are neither recycled nor trashed but stockpiled in drawers. And they are good computers: recent smartphone processors rival an Intel Core i3, and the paper’s opening figure shows phone specs meeting or exceeding the vCPU/memory envelopes of AWS’s t4g microservice instances. Cars get resold and driven into the ground; computers, uniquely, are abandoned while functional.

The missing piece is a metric.

What people did before, and where it broke

Mobile SoCs for HPC. Rajovic et al. showed mobile chips are performant and efficient enough for many HPC codes — but used new chips, which pays the manufacturing carbon this paper wants to avoid. Shahrad and Wentzlaff proposed a decommissioned-phone server as a design sketch, without implementation or carbon quantification; Büsching et al. ran LINPACK over WiFi on six phones. Follow-on work by Gupta et al. (the ACT line) built carbon-modeling tools and suggested phones-as-SSD-storage — kindred spirit, different workload.

Phones as sensors. IoT deployments learned a hard lesson the paper inherits: mobile operating systems assume a human is present, and removing the human exposes “a very long tail of deployment-ending issues.” Hence the prototype’s first move is discarding Android entirely.

Life-cycle assessments. Vendors (Dell, Apple) publish LCAs giving total lifetime carbon per device. LCAs are the raw material here, but a total is not a rate: it cannot say whether five more months of service is worth it. The paper’s metric turns LCA totals into an amortization argument.

The idea

Make longevity a number. CCI’s arithmetic is deliberately simple: a device’s carbon is a large fixed manufacturing term plus a usage term that accrues with energy drawn; its value is the work it performs. Dividing carbon by work makes the fixed term amortizable — every additional month of service lowers the average — and lets old and new devices compete on one axis. It also redefines “end of life”: a device should retire not when a newer one exists, but when the newer device’s extra work per unit carbon overtakes the old one’s already-paid manufacturing. Under a hypothetical 100% carbon-free grid, operational carbon goes to zero and embodied carbon becomes the whole game — reuse wins by construction.

Phones are the best junk. Among the candidates, old servers lose on power draw (a 15-year ProLiant needs 20 units to match the R740 and burns server-class watts); old laptops are decent but scarce. Phones combine energy-efficient SoCs, integrated storage and WiFi, colossal supply, and — decisive twice over — a battery. The battery is a built-in UPS (checking the datacenter backup-power requirement for free) and an energy buffer that enables carbon-aware smart charging. Even thermals cooperate: phones are engineered never to burn a hand, so they self-throttle; the paper’s styrofoam-box stress test measured ~2.6 W of thermal output per device at 100% load, meaning a 256-phone cloudlet (~666 W thermal) needs two commodity server fans.

Aggregate to server scale. One R740’s light-medium workload (Dell’s own LCA operating profile) can be matched by 54 Pixel 3As at 84 W — versus the server’s 309 W — or 256 Nexus 4s at 456 W. The startling result: even the higher-powered Nexus 4 cluster beats the new server’s CCI on two of three benchmarks at any lifetime, and on the third (SGEMM, where the compute gap is 256×) whenever the server would serve fewer than 45 months. Embodied carbon is that heavy.

Prove it on real software. The prototype runs DeathStarBench — complete social-network and hotel-reservation applications built from dozens of Docker microservices — on ten phones. Mobile OS assumptions are the real enemy: Android goes, Ubuntu Touch with a patched kernel (BTRFS for Docker) comes in.

hardware: 10 × Pixel 3A(/XL), ~$65 each
OS: Ubuntu Touch + patched kernel, Docker-capable

Why it works

• Manufacturing dominates lifetime carbon for consumer devices (70–85%) → extending service life amortizes an already-spent cost → reuse supplants manufacturing, the single largest carbon line item.
• Microservices right-sized the cloud → individual services need t4g-class cores and gigabytes, not sockets and terabytes → a 2019 phone clears the bar, and aggregation covers the rest.
• Phones are self-contained infrastructure → battery = UPS + smart charging buffer; thermal self-limiting = nearly free cooling; WiFi = built-in networking → the peripherals a junk-server needs are already in the junk.
• The grid is decarbonizing → operational carbon trends toward zero → embodied carbon becomes the dominant term everywhere, widening reuse’s advantage over time (the 100%-solar analysis shows it).
• CCI prices all of this in one number → device retirement becomes a calculation (45 months for the SGEMM crossover) instead of a fashion cycle.

What it costs

• Performance ceilings. Compute-dense kernels (SGEMM’s 256× gap) still favor new silicon; the Nexus 4 cluster loses that benchmark beyond 45-month server lifetimes.
• Operational complexity. Hundreds of devices, USB charging, battery replacements every ~2 years, custom OS images, and a 20% overhead of phones doing networking/management duty.
• WiFi networking. 18.5 Mbit/s per device in the tree topology is fine for the benchmarked services, far from datacenter fabric.
• Estimation uncertainty. EC2 power and embodied carbon are third-party estimates; phone LCAs vary by vendor; CCI inherits LCA quality — a limitation the paper turns into a call for disclosure.
• Space. Phone clusters are volumetrically inefficient (the 50 MW sketch leaves 75% of rack space empty), slightly worsening facility overheads like lighting and cooling.

The evaluation

Three layers of evidence. Device level: Geekbench and power characterization across the five platforms under Dell’s own light-medium load profile. Cluster level: CCI computed for R740-equivalent cloudlets (17 ThinkPads / 20 ProLiants / 54 Pixel 3As / 256 Nexus 4s) across three benchmarks and three energy regimes (California mix, 24/7 solar, theoretical zero-carbon). System level: the ten-phone prototype versus EC2 on DeathStarBench.