Skip to content

Volunteer Computing

Abstract

In 1999, before “cloud” meant anything, UC Berkeley asked a radical question: what if the world’s idle screensavers were a supercomputer? SETI@home (May 17, 1999) shipped a program that downloaded chunks of Arecibo radio data, hunted them for alien signals while your PC sat idle, and sent the answers home. It became the largest computation in history (Guinness, 2008) and enrolled millions of people who mostly wanted the animated screensaver. The model (volunteer distributed computing) generalized into BOINC, the platform behind dozens of science projects, and reached its apotheosis in Folding@home, which during the COVID-19 pandemic in 2020 crossed an exaFLOP of aggregate power, becoming, for a moment, faster than the top 500 supercomputers combined, assembled entirely from strangers’ gaming rigs. It never found any aliens. It also proved that scientific supercomputing could be crowdsourced, and rehearsed, a decade early, every idea that GPU cloud computing would later monetize.

The Screensaver Supercomputer

The intellectual precursor was GIMPS (Great Internet Mersenne Prime Search, 1996), which found record primes by farming the arithmetic to volunteers; distributed.net (1997) cracked cryptographic challenges the same way. But the breakthrough in reach was SETI@home. The idea came to David Gedye, an Australian software developer at Paul Allen’s start-up Starwave in Seattle, at his 35th birthday party on 11 December 1994, during a conversation about whether anything could ever grip the public the way the Moon landing had. The astronomer Woodruff Sullivan put him in touch with Dan Werthimer, who ran a SETI programme at Berkeley, and the distributed-systems specialist David Anderson joined them. Starwave gave $10,000 in 1997, Paramount and the Planetary Society $50,000 each, and Sun Microsystems contributed too. SETI@home launched on May 17, 1999 from Berkeley’s Space Sciences Laboratory; the client was a 700-kilobyte download, about five minutes on a 28.8 kbit/s modem. The genius was packaging: the client ran as a screensaver showing a pretty waterfall plot of the radio spectrum it was analyzing, so participation was both effortless and visible, a status symbol, a conversation piece, a way for a home PC to matter. Users churned through data from the Arecibo telescope in Puerto Rico, each machine sifting a slice of sky for the signal shapes (spikes, Gaussian rises and falls, triplets, pulses) that would distinguish a transmitter from cosmic noise.

A week after launch 298,000 people were taking part, the millionth arrived in August 1999 and the two millionth less than a year later; in 2013 some 145,000 computers in 233 countries were still active. Millions signed up across 200+ countries. The aggregate throughput dwarfed the era’s fastest single machines, and Guinness certified it in 2008 as the largest computation in history. It found no confirmed extraterrestrial signal in twenty-one years (a 2004 candidate, SHGb02+14a, went nowhere), and on March 31, 2020 Berkeley stopped distributing new work, citing diminishing returns and a backlog of analysis. Its real discovery was social: that ordinary people would donate their electricity and CPU cycles to open science if you made it easy and let them watch.

BOINC: The Platform

Anderson generalized the machinery into BOINC (Berkeley Open Infrastructure for Network Computing, 2002), a free platform that handled the hard parts common to every such project: distributing work units, validating returned results against fraud and hardware error (send each unit to multiple volunteers, compare), crediting contributors, and scheduling across a volunteer’s chosen projects. SETI@home moved onto it in 2004 or 2005 (accounts differ), and BOINC became the substrate for a research commons: Rosetta@home (protein structure), Einstein@home (gravitational-wave and pulsar searches, which really did discover new pulsars), Climateprediction.net (climate-model ensembles; see Weather and Climate Modeling), World Community Grid (disease research, backed by IBM), and genomics projects. The validation-by-redundancy design is a small classic of distributed systems engineering: assume every node is unreliable and possibly lying, and still produce trustworthy science.

Folding@home and the Exascale Moment

Folding@home, launched October 1, 2000 by Vijay Pande at Stanford, aimed the same idea at protein folding, simulating how proteins contort into their functional shapes, the molecular events behind Alzheimer’s, cancer, and viral infection. Its most famous engine was unexpected: the PlayStation 3. From 2007 to 2012 Sony shipped a Folding@home client for the console’s powerful Cell processor; over 15 million PS3 users contributed more than 100 million compute-hours, at times the largest single source of the project’s power (see The Video Game Console Wars for the Cell’s story).

Then came the pandemic. As COVID-19 spread in March 2020, hundreds of thousands of new volunteers joined to simulate the SARS-CoV-2 spike protein and hunt druggable pockets. Aggregate power rocketed past 1.5 exaFLOPS (x86-equivalent; 768 native petaFLOPS) on March 25, 2020 and 2.43 exaFLOPS by mid-April, making Folding@home, by that measure, the world’s first exascale computing system, and briefly more powerful than the entire Top500 supercomputer list combined, two years before the first “official” exascale machine (Frontier, 2022) existed. The distributed collective had beaten the purpose-built giants to the milestone (see The Supercomputer Era). The project has produced 200+ peer-reviewed papers; it is now led by Greg Bowman at Penn.

⚠️ Dead End: The Screensaver Model, Overtaken

Volunteer computing’s decline is a story of being right too early and then being out-economized. Three forces eroded it: GPUs and the cloud made raw FLOPS a commodity you could rent by the hour (the GPU revolution, cloud computing), for a funded lab, buying AWS time beats begging for screensaver cycles; AlphaFold (2020–21) largely solved static protein-structure prediction with a trained neural network, doing in minutes what Rosetta@home had attacked with volunteer cycles for years, although simulating how proteins move and fold, Folding@home’s task, remained (see AlphaFold); and cryptocurrency turned the very idea of donating spare compute into a market, then poisoned it, the same “use your idle GPU” pitch became mining, and worse, cryptojacking malware that stole exactly the cycles volunteer projects had asked for politely. The altruistic screensaver could not compete with a paying one. Yet the core proof endures and was arguably vindicated: idle consumer hardware, aggregated, is a supercomputer, and the peer-to-peer, radically-decentralized compute idea keeps being reinvented (in blockchain networks, in distributed AI-training experiments) usually by people who have forgotten that a screensaver hunting for aliens got there first, in 1999.

Fun Fact

In September 2007 Folding@home, carried mostly by PlayStation 3 consoles, became the first computing system of any kind to sustain one petaFLOPS, eight months before IBM’s Roadrunner reached the same mark as a single machine in May 2008. The most powerful computer on Earth had no building and no budget line; it was game consoles and home PCs left switched on overnight.

📚 Sources