The long running race to calculate more digits of pi has reached a new milestone. A team at StorageReview has reclaimed the world record by computing 314 trillion digits of pi using a single on-premise server that ran continuously for nearly four months.
What makes the achievement notable is not only the scale of the calculation, but how it was achieved. Instead of relying on cloud clusters or distributed storage systems, the team demonstrated that raw local storage bandwidth can be the decisive factor in ultra-large pi computations.
Pi calculations were once a straightforward way to benchmark floating-point CPU performance. As record attempts pushed into the trillions of digits, the challenge shifted. Memory capacity, I/O architecture, and storage throughput became just as critical as raw compute power.
The entire run was completed on a single 2U server, avoiding the complexity of clustered systems used in earlier records. Previous milestones, such as Google’s 100-trillion-digit calculation or the more recent 300-trillion run using shared storage, depended heavily on distributed infrastructure. StorageReview instead focused on maximizing performance within one machine.
The system was a Dell PowerEdge R7725 fitted with two AMD EPYC processors, each offering 192 cores, for a total of 384 cores. It was paired with 1.5 terabytes of DDR5 memory.
Where the setup truly excelled was storage. The server housed a 40-drive array of Micron 6550 ION SSDs, each with 61.44 terabytes of capacity, delivering roughly 2.5 petabytes of local storage. Unlike earlier designs, this generation of Dell servers connects the drives directly to the CPUs’ PCIe lanes, eliminating a PCIe switch in the backplane. Even with only a few lanes per drive, the configuration sustained around 280 gigabytes per second of read and write bandwidth.
Additional optimizations helped sustain the four-month computation. The scratch storage layout was tuned specifically for the access patterns generated by the y-cruncher software at extreme digit counts. Cooling was upgraded to a CoolIT liquid system, allowing the EPYC processors to hold higher steady-state clocks under continuous load. The operating system was also switched from Windows Server to Ubuntu 24.04.2, which delivered measurable gains in I/O performance.
Despite the scale of the hardware, total power draw averaged about 1,600 watts. While high in absolute terms, it is considered efficient given the amount of computation and storage activity involved.
Beyond the headline number, the record underscores a broader lesson. At extreme scales, calculating pi is no longer just about CPU speed. Storage architecture and bandwidth can determine whether a record attempt succeeds. By reaching 314 trillion digits on a single, carefully tuned server, StorageReview has shown that local, high-performance storage can rival or surpass far more complex distributed approaches.
