Italian
StorageReview Sets New Pi Record: 314 Trillion Digits Calculated
StorageReview has reclaimed the prestigious computational crown by calculating π to an astonishing 314 trillion digits, shattering the previous world record. The modern race for π precision has evolved from initial cloud experiments into a full-scale showcase of infrastructure prowess and engineering ingenuity.
Back in 2022, Google Cloud made headlines by computing π to 100 trillion digits. Leveraging a massive fleet of cloud instances to run the y-cruncher program, the project consumed tens of petabytes of I/O data. At the time, this milestone stood as the definitive benchmark for the upper limits of what traditional infrastructure could achieve.
The focus then shifted to on-premises lab environments. In early 2024, we upgraded our system to set a new record of 105 trillion digits, supported by nearly 1 petabyte of Solidigm QLC SSDs. This achievement established a new scale benchmark, proving that a single on-premises machine could deliver exceptional efficiency. Just months later, we pushed the boundaries further to 202 trillion digits. This breakthrough validated that high-density flash storage, combined with meticulous system tuning, could outperform hyperscale cloud infrastructure for this highly demanding specific workload.
Naturally, every record invites a challenge. Linus Media Group and KIOXIA subsequently claimed the title by computing π to 300 trillion digits. Their setup relied on a large Weka shared-storage cluster equipped with 2PB of flash storage. While this demonstrated the potential of storage-heavy traditional infrastructure, it came with significant trade-offs: a full rack of hardware, substantial power consumption, and complex cooling requirements. We were determined not to let this record stand unchallenged.
Today, StorageReview is proud to announce our new victory: π calculated to 314 trillion digits. The feat was achieved using a single 2U Dell PowerEdge R7725 server, fitted with dual AMD EPYC 192-core CPUs and forty 61.44TB Micron 6550 Ion SSDs. We completed the system build and tuning in July, kicking off the calculation run on July 31, 2025. Coincidentally, the run concluded on the second day of SC25, making this new high-performance computing (HPC) record all the more timely.
Scaling y-cruncher to 314 Trillion Digits
Once the calculation scale surpasses hundreds of trillions of digits, y-cruncher transforms from a traditional benchmark into a grueling long-haul infrastructure stress test. The program's core logic remains straightforward, but its interaction with hardware at this scale becomes the decisive factor. The entire operation hinges on the system's ability to keep thousands of multi-precision operations running smoothly—without stalling the CPUs or overwhelming the storage layer. In fact, this new record was truly won in the storage layer.
We deployed 40 Micron 6550 Ion Gen5 NVMe SSDs, with 34 allocated exclusively to run y-cruncher. This SSD pool provides approximately 2.1 petabytes of usable space, giving y-cruncher the ample storage it needs to compute and process 314 trillion digits of π. The remaining 6 SSDs are configured into a software RAID10 volume, specifically used to store the final 314-trillion-digit results.
Design improvements between the 16th and 17th generations of Dell PowerEdge servers also played a key role in boosting performance for this 314-trillion-digit record run. Our previous 202-trillion-digit attempt used the 24-bay Dell PowerEdge R760, which featured a PCIe switch on the drive backplane—trading drive density for enhanced performance. In contrast, the 17th-generation PowerEdge servers, such as the Intel-based R770 and AMD-based R7725, feature backplanes with direct-connect-only design, offering 2 or 4 PCIe lanes per bay.
The PowerEdge R7725, equipped with its 40-bay Gen5 E3.S backplane, provides 2 PCIe lanes per SSD slot. While this may seem like a potential performance trade-off on paper, the platform can still deliver impressive speeds: up to 280GB/s of simultaneous read and write performance when all 40 bays are fully utilized.
Using the internal y-cruncher storage benchmark, we measured the storage performance of each platform across its respective configuration. Across all workloads, we observed storage performance improvements ranging from 72% to 383%, with well-balanced read and write metrics—solidifying the 17th-generation PowerEdge server's superiority for high-scale storage-intensive computing.
| Metric | 202T System (old record) | 314T System (new record) | Percent Difference (314T vs 202T) |
|---|---|---|---|
| Sequential Write | 47.0 GiB/s | 107 GiB/s | +127.7% |
| Sequential Read | 56.7 GiB/s | 127 GiB/s | +124.0% |
| Threshold Strided Write | 62.2 GiB/s | 107 GiB/s | +72.0% |
| Threshold Strided Read | 20.9 GiB/s | 101 GiB/s | +383.3% |
The Dell PowerEdge R7725 is far more than just a storage powerhouse; it shines brightly as a dual-socket AMD Turin platform with exceptional compute potential. We equipped our system with 192-core AMD EPYC 9965 processors, delivering a total of 384 cores. To unlock this performance, we upgraded the stock air-cooling solution to liquid-cooled CoolIT SP5 cold plates, chilled by a CoolIT AHx10 Liquid-to-Air CDU (Coolant Distribution Unit).
This strategic cooling upgrade delivered three key advantages: it allowed the CPUs to maintain sustained high clock speeds, kept the server's chassis fans running at a remarkably low 30% PWM (Pulse Width Modulation), and ensured the system's average power consumption remained steady at around 1,600W.
Software & System Optimization
For the software layer, we departed from past practices by choosing Ubuntu 24.04.2 LTS Server instead of Windows Server. This decision proved instrumental in maximizing system stability and unlocking significant gains in workload performance.
Before initiating the record run, we conducted rigorous test iterations and fine-tuned the configuration. A critical step was reserving 4 out of the 384 available cores for critical background system operations, ensuring the main computation threads remained unimpeded. The result? We didn't just beat the previous pi record—we obliterated it across every measurable metric. Our run stands unparalleled in performance, power efficiency, and most impressively, in its flawless reliability. We are uniquely proud to hold the distinction of being the only large-scale pi world-record attempt completed with zero downtime; the calculation ran seamlessly from start to finish without ever needing to be resumed.
Record-Breaking Power Efficiency
A core principle behind each of StorageReview's pi record attempts is to minimize system complexity and optimize for energy efficiency. The previous 300T record relied on a distributed storage cluster with high-speed networking, which inherently demanded larger power and cooling budgets.
We took a fundamentally different approach. By focusing on extreme storage density, we consolidated both swap space and persistent output storage onto a single 2U server. This architectural choice was pivotal in drastically reducing our overall power and cooling footprint.
Over the course of the full 314-trillion-digit calculation, our Dell PowerEdge R7725 consumed a total of 4,304.662 kWh. This equates to an extraordinary 13.70 kWh per trillion digits—placing our solution among the most energy-efficient large-scale pi computations ever performed. The stark contrast between our efficient, single-server approach and the power-hungry distributed cluster method is immediately clear, as highlighted in the comparison table below.
| Run | Total kWh | Cost @ $0.12/kWh | Cost @ $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 crucial to highlight that during our 314-trillion-digit calculation, we utilized SSDs in a JBOD (Just a Bunch of Disks) configuration, which does not include built-in data resiliency. This decision was driven by two core priorities: optimizing power consumption and maximizing overall system performance. However, it also sparked an important conversation about tailoring storage solutions to the specific demands of the workload at hand. Every workload is unique—some can be restarted with minimal impact on operations and therefore do not require the same high level of fault tolerance as mission-critical tasks. In our case, we focused on safeguarding the final data output through a traditional software RAID setup, ensuring the integrity of the record-breaking π digits without compromising the system’s efficiency.
110 Days of Uninterrupted Runtime
Despite computing more digits than any previous π record attempt, our wall-clock runtime was significantly shorter than the prior benchmark. The previous 300-trillion-digit record required approximately 225 total days to complete—equating to 175 actual compute days when excluding downtime. In stark contrast, our 314-trillion-digit run achieved a full 110 days of uninterrupted operation. This exceptional stability can be attributed to four key factors: a robust and stable operating system, a minimized background workload to avoid resource contention, a balanced NUMA (Non-Uniform Memory Access) topology for optimal CPU-memory communication, and a scratch array specifically engineered to handle the unique data access patterns generated by y-cruncher at this unprecedented scale.
Technical Highlights
- Total 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 and Algorithms: y-cruncher v0.8.6.9545, Chudnovsky
- SSD Wear per SMART: 7.3PB written per Drive or 249.11PB across the 34 SSDs used for swap
- Logical Largest Checkpoint: 850,538,385,064,992 (774 TiB)
- Logical Peak Disk Usage: 1,605,960,520,636,440 (1.43 PiB)
- Logical Disk Bytes Read: 148,356,635,606,263,504 (132 PiB)
- Logical Disk Bytes Written: 126,658,805,195,776,600 (112 PiB)
- Start Date: Thu Jul 31 17:16:41 2025
- End Date: Tue Nov 18 05:57:08 2025
- pi: 8793223.144 seconds, 101.773 Days
- Total Computation Time: 9274878.580 seconds
- Start-to-End Wall Time: 9463226.454 seconds
Closing Thoughts
For decades, extreme π calculations have served as a showcase for computing prowess, each era's "big iron" taking center stage. Early records relied on high-performance desktops and external storage arrays; the focus then shifted to on-premises enterprise gear. More recently, the race migrated to the cloud, where initiatives like Google’s 100-trillion-digit run demonstrated that brute force—armed with thousands of instances and massive I/O throughput—could secure a record. We then saw the emergence of large shared-storage clusters, prioritizing raw parallelism over simplicity, but at the cost of substantial power consumption and complex cooling demands.
Our journey has taken the opposite path. Across successive record-setting runs, we’ve treated the y-cruncher workload as a critical, long-haul HPC task—not a one-off stunt. The 105T and 202T campaigns were instrumental in identifying core bottlenecks: optimizing scratch storage capacity, ensuring a steady CPU workload without overwhelming the I/O layer, and hardening the system to deliver reliable performance over months. The 314T run is the culmination of that hard-won expertise. This isn’t merely a larger number; it represents a more mature, production-grade design.
The data speaks for itself. We shattered the 300-trillion-digit barrier on a single 2U Dell PowerEdge R7725, outfitted with 40 Micron 6550 Ion SSDs and dual 192-core AMD EPYC CPUs. The system ran flawlessly for 110 consecutive days, completing the calculation without a single interruption or restart. Storage throughput more than doubled compared to our 202T platform, yet the server maintained a modest average power draw of 1,600W, with a total energy consumption of just 4,305 kWh. This translates to a remarkably efficient 13.70 kWh per trillion digits—a fraction of the estimated power used by the previous 300T cluster. In short, we achieved more work with fewer nodes, less complexity, and lower energy costs.
This record’s significance extends far beyond bragging rights. If a single commercial 2U server can sustain a workload of this magnitude with such reliability and efficiency, the same architectural principles translate directly to real-world scientific computing. Long-running climate models, high-precision physics simulations, genomics pipelines, and large-scale AI training jobs all depend on the same foundational elements: balanced I/O performance, predictable thermal management, stable firmware, and an architecture built to operate continuously for months. Our platform has now proven it can deliver on these fundamentals under the most demanding conditions.
Yes, StorageReview has reclaimed the π crown with 314 trillion digits. More importantly, we’ve raised the bar for what "excellence" means in large-scale numerical computing on physical hardware. For any challengers seeking to break this record, we issue a clear challenge: aim for the complete package—more digits, lower power, shorter runtime, and the same zero-downtime reliability. Until then, this setup stands as the definitive benchmark for efficiency and robustness in the field.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Dettagli di contatto
Beijing Qianxing Jietong Technology Co., Ltd.
Persona di contatto: Ms. Sandy Yang
Telefono: 13426366826