StorageReview has reclaimed a preferred computational crown with a brand new pi document clear up to 314 trillion digits. The trendy π race has gone from cloud experiments to full-blown infrastructure flex. In 2022, Google Cloud pushed π to 100 trillion digits, working y-cruncher throughout a large fleet of cloud situations and chewing via tens of petabytes of I/O within the course of. That mark stood because the headline quantity for “how far is feasible” with conventional infrastructure.
The motion then shifted into the lab. In early 2024, we upgraded our document to 105 trillion digits on a system supported by practically a petabyte of Solidigm QLC SSDs. This achievement set a brand new benchmark for scale and demonstrated how effectively a single on-premises machine may function. A number of months later, we did it once more, this time reaching 202 trillion digits. This validated that high-density flash and cautious tuning may outperform hyperscale infrastructure for this particular, demanding workload.
In fact, data invite challengers. Linus Media Group and KIOXIA grabbed the crown with a 300 trillion-digit run, powered by a big Weka shared-storage cluster with 2PB of flash. That effort confirmed what storage-heavy conventional infrastructure may do, albeit with a rack of {hardware}, a big energy invoice, and cooling complexities. We couldn’t idly stand by and let that document stand!
StorageReview has now pushed π to 314 trillion digits, utilizing a single 2U Dell PowerEdge R7725 server geared up with twin AMD EPYC 192-Core CPUs and forty 61.44TB Micron 6550 Ion SSDs. With the system construct and tuning happening in July, we kicked off our run on July 31, 2025. As luck would have it, our pi run wrapped throughout our second day of SC25, becoming for a brand new HPC document.
Scaling y-cruncher to 314 Trillion Digits
When you cross into the a whole bunch of trillions of digits, y-cruncher behaves much less like a conventional benchmark and extra like a long-haul infrastructure stress check. The applying itself is simple, however the way in which it interacts with {hardware} at this scale turns into the figuring out issue. Every thing comes right down to how effectively the system can preserve 1000’s of multi-precision operations transferring with out stalling the CPUs or overwhelming the storage layer. The storage layer, particularly, is the place this document was truly gained. We deployed 40 Micron 6550 Ion Gen5 NVMe SSDs, 34 of which had been allotted to y-cruncher. This SSD pool delivered about 2.1 petabytes, giving y-cruncher sufficient area to carry out the 314T run. The remaining 6 SSDs had been used to construct a software program RAID10 quantity, which we used to document the 314T digits of pi.
Design modifications between the Sixteenth- and Seventeenth-generation Dell PowerEdge servers additionally contributed to bettering the efficiency of our most up-to-date 314T document run. In our earlier 202T document, we leveraged the 24-bay Dell PowerEdge R760, which used a PCIe swap on the drive backplane, buying and selling drive density for efficiency. On the Seventeenth-generation Dell PowerEdge servers, such because the Intel-based R770 or AMD-based R7725, the backplanes reverted to direct-connect-only, with 2 or 4 PCIe lanes per bay. Within the PowerEdge R7725 with its 40-bay Gen5 E3.S backplane, every SSD will get 2 PCIe lanes. Whereas which will seem to be a efficiency hit, the platform remains to be able to reaching as much as 280GB/s learn and write when all 40 bays are hit concurrently.
Utilizing the interior y-cruncher storage benchmark, we recorded the storage efficiency of every platform within the configuration used for every run. Throughout every workload, we noticed storage efficiency will increase starting from 72% to 383%, with balanced learn and write metrics.
| Metric | 202T System (previous document) | 314T System (new document) | % Distinction (314T vs 202T) |
|---|---|---|---|
| Sequential Write | 47.0 GiB/s | 107 GiB/s | +127.7% |
| Sequential Learn | 56.7 GiB/s | 127 GiB/s | +124.0% |
| Threshold Strided Write | 62.2 GiB/s | 107 GiB/s | +72.0% |
| Threshold Strided Learn | 20.9 GiB/s | 101 GiB/s | +383.3% |
The Dell PowerEdge R7725 isn’t only a storage field; it additionally provides vital compute potential as a dual-socket AMD Turin platform. We leveraged 192-core AMD EPYC 9965 CPUs, totaling 384 cores. We additionally changed the inventory air-cooled heatsinks with liquid-cooled CoolIT SP5 chilly plates, which had been cooled by a CoolIT AHx10 Liquid-to-Air CDU. This mixture saved the CPUs in greater sustained clock speeds, chassis followers working at round 30% PWM, and our common system energy consumption hovering round 1,600W.
On the software program facet, the platform ran Ubuntu 24.04.2 LTS Server as an alternative of Home windows Server as in previous runs. This maximized system stability and delivered vital good points in workload efficiency. We carried out quite a few check iterations earlier than committing to kicking off our run, together with reserving 4 out of 384 cores for background system operations. The consequence wasn’t simply beating the prevailing pi document; we obliterated it throughout quite a few metrics. Nothing comes near our run by way of efficiency, energy consumption, and most impressively, reliability. We’re additionally the one large-scale pi world-record run with out a second of downtime. From begin to end, the run by no means needed to be resumed.
File-Setting Energy Effectivity
The method StorageReview has taken for every of our pi document runs has been to cut back complexity and run the benchmark utilizing as little energy as essential. The earlier 300T document, which leveraged a distributed storage cluster and a high-speed community, got here at the price of bigger energy and cooling necessities. We took a unique path, specializing in storage density to make use of a single 2U server for each swap and output storage. This performed a major position in lowering our general energy and cooling footprint. Our Dell PowerEdge R7725 consumed simply 4,304.662 kWh over the course of the 314T run, which works out to only 13.70 kWh per trillion digits. This makes it one of the energy-efficient large-scale pi computations. When evaluating the 2 approaches, the distinction turns into instantly obvious, as highlighted within the desk beneath.
| Run | Complete kWh | Price @ $0.12/kWh | Price @ $0.20/kWh |
|---|---|---|---|
| 300T Weka Cluster Run | 33,600 kWh (est.) | $4,032 | $6,720 |
| 314T Single-Server Run | 4,304.662 kWh | $517 | $861 |
It is very important observe that in the course of the calculation, our 314T run leveraged SSDs in a JBOD configuration with out knowledge resiliency. Energy consumption and general system efficiency had been the components that led to that call, but it surely additionally sparked a dialogue about designing the storage resolution across the workload. Each workload is completely different, and a few that may be restarted with minimal influence on manufacturing might not require the identical stage of fault tolerance. In our case, we centered on defending the information output with conventional software program RAID.
110 Days Complete Runtime
Regardless of computing extra digits than any run earlier than it, the wall-clock time was considerably decrease than the earlier document, which required roughly 225 days to finish (175 compute days excluding downtime). The uninterrupted 110-day window is the results of a secure OS, a minimized background load, a balanced NUMA topology, and a scratch array engineered for the sample y-cruncher generates at this scale.
Technical Highlights
- Complete Digits Calculated: 314,000,000,000,000
- {Hardware} Used: Dell PowerEdge R7725 with 2x AMD EPYC 9965 CPUs, 1.5TB DDR5 DRAM, 40x Micron 61.44TB 6550 Ion
- Software program and Algorithms: y-cruncher v0.8.6.9545, Chudnovsky
- SSD Put on per SMART: 7.3PB written per Drive or 249.11PB throughout the 34 SSDs used for swap
- Logical Largest Checkpoint: 850,538,385,064,992 (774 TiB)
- Logical Peak Disk Utilization: 1,605,960,520,636,440 (1.43 PiB)
- Logical Disk Bytes Learn: 148,356,635,606,263,504 (132 PiB)
- Logical Disk Bytes Written: 126,658,805,195,776,600 (112 PiB)
- Begin Date: Thu Jul 31 17:16:41 2025
- Finish Date: Tue Nov 18 05:57:08 2025
- pi: 8793223.144 seconds, 101.773 Days
- Complete Computation Time: 9274878.580 seconds
- Begin-to-Finish Wall Time: 9463226.454 seconds
Closing Ideas
For many years, excessive pi runs had been a technique to exhibit no matter counted as “huge iron” on the time. Early data leaned on high-performance desktops and externally hooked up storage, then moved in the direction of native enterprise gear. Extra not too long ago, the race moved into the cloud, the place jobs like Google’s 100 trillion digit run proved that you could possibly brute power your technique to a document with sufficient situations and sufficient I/O. Then we noticed massive shared storage clusters step in, buying and selling simplicity for uncooked parallelism and an amazing energy and cooling invoice.
Our method has gone in the wrong way. Over a number of document runs, we’ve handled y-cruncher as a severe HPC workload, not a one-off stunt. The 105T and 202T options helped us establish the true bottlenecks, dimension and tune scratch storage, preserve CPUs fed with out thrashing the I/O layer, and harden a system so a months-long job truly finishes. The 314T run is the results of that have. It’s not only a greater quantity. It’s a extra mature design.
The metrics reaffirm that story. We pushed previous 300 trillion digits on a single 2U Dell PowerEdge R7725 with 40 Micron 6550 Ion SSDs and twin 192-core AMD EPYC CPUs, saved the system on-line for 110 straight days, and by no means needed to resume from any failure. Storage throughput greater than doubled in comparison with our 202T platform, but the server averaged about 1,600W and consumed 4,305 kWh in whole. That works out to 13.70 kWh per trillion digits, which is a fraction of the estimated power utilized by the earlier 300T cluster: fewer nodes, much less complexity, much less power, extra work completed.
That’s the reason this document issues past bragging rights: if one business 2U server can maintain a y-cruncher run of this dimension, with that stage of reliability and effectivity, the identical design patterns map on to manufacturing science. Lengthy-running local weather fashions, physics simulations, genomics pipelines, and AI coaching jobs all reside or die on the identical fundamentals: balanced I/O, predictable thermals, secure firmware, and an structure that may keep upright for months at a time. This platform has now confirmed it will possibly do exactly that beneath circumstances that go away no room for error.
Sure, StorageReview has reclaimed the pi crown with 314 trillion digits. Extra importantly, we’ve set the bar for what “good” means in large-scale numerical computing on actual {hardware}. If somebody desires to take the document, we want to see them take the entire thing: extra digits, much less energy, shorter wall time, and the identical zero-downtime reliability. Till then, that is the benchmark for effectivity.
