Numinous

Numinous › Invention › Fact
Invention

In 2019, Emma Haruka Iwao calculated pi to 31,415,926,535,897 digits — on rented cloud servers, not a supercomputer.

3 sources verified 3 min read
In 2019, Emma Haruka Iwao calculated pi to 31,415,926,535,897 digits — on rented cloud servers, not a supercomputer.ILLUSTRATION · AI

In 2019, Google developer Emma Haruka Iwao set a world record by computing π to 31,415,926,535,897 digits on rented cloud servers rather than a dedicated supercomputer.

On March 14, 2019 — Pi Day — Emma Haruka Iwao announced that she had calculated Archimedes' constant to 31,415,926,535,897 digits. That figure is not arbitrary: read the first digits and you'll see 31.4 trillion, a number chosen because it echoes π itself. It was, at the time, the most accurate value of π ever computed, and it was done on infrastructure anyone with a credit card could rent.

The real break with tradition wasn't the digit count. It was the machine — or rather, the lack of one. Every previous record had leaned on a single, powerful computer or a purpose-built rig. Iwao's run was the first to claim the record using a commercial public cloud.

How she calculated 31.4 trillion digits of pi

Iwao, a developer advocate at Google, laid out the mechanics herself in a Google Cloud Blog post titled "Pi in the sky." The computation used y-cruncher, the multi-threaded program written by Alexander Yee that has become the standard tool for extreme π calculations. Rather than running it on one machine, she distributed the job across a cluster of 25 Google Compute Engine virtual machines.

As Richard Speed reported in The Register, those were Intel-powered GCE instances, and they ran for 121 days straight — four months of uninterrupted arithmetic. The scale of data involved is where the problem stops being abstract. Computing tens of trillions of digits means shuffling enormous intermediate results that no single computer's memory can hold, so the calculation spills onto storage. Iwao's job touched roughly 170 terabytes of data, about the same order as the entire printed collection of the Library of Congress, moving through disks over and over as the algorithm ground forward.

Getting a number is easy; proving it's correct is the hard part. Yee independently verified the result using the Bailey–Borwein–Plouffe (BBP) formula and Bellard's formula — methods that can compute specific digits of π far out in the sequence without recomputing everything before them. If the independently calculated digits match the end of Iwao's run, the whole calculation is almost certainly right. They matched.

Why the cloud mattered more than the digits

The contrast with the prior record makes the point. As Jon Porter noted in The Verge, the previous holder, Peter Trueb, had done his 2016 calculation on a single computer. Iwao's approach swapped the one-big-machine model for rented, networked, commodity hardware — the same kind of virtual machines businesses use to run websites and databases. It demonstrated that a task once reserved for institutional supercomputers could be assembled on demand and torn down when finished.

That's the quiet significance behind the stunt. A 31-trillion-digit value of π has no practical use; NASA has said that 15 or 16 digits are enough to navigate spacecraft across the solar system. The record was never about needing more digits. It was a stress test of what cloud infrastructure could do, demonstrated in public, with a result anyone could check.

There's one vivid detail that captures the fragility of the whole enterprise. A 121-day computation across 25 machines is a long time for nothing to fail. A single corrupted disk, a dropped network connection, or a crashed instance could threaten four months of work — which is exactly why cloud systems that can snapshot, restart, and route around failure made the attempt feasible in the first place. The achievement is partly mathematical and partly a demonstration that distributed systems can stay coherent that long.

The seam still worth arguing over is what "record" even means now. Iwao's 31.4 trillion was itself surpassed within a couple of years, and each new mark depends on who can rent — or build — more computing time. When the limit is budget and patience rather than ingenuity, is it still a record of human achievement, or just a receipt for compute hours? That question doesn't have a verified answer.

SourcesEvery claim traced
Advertisement
Keep reading

More facts