StorageReview가 새로운 Pi 기록을 세웠습니다: Dell PowerEdge R7725에서 314T 자릿수

StorageReview has reclaimed a popular computational crown with a new pi record solve to 314 trillion digits. The modern π race has gone from cloud experiments to full-blown infrastructure flex. In 2022, Google Cloud pushed π to 100 trillion digits, running y-cruncher across a massive fleet of cloud instances and chewing through tens of petabytes of I/O in the process. That mark stood as the headline number for “how far is possible” with traditional infrastructure. The action then shifted into the lab. In early 2024, we upgraded our record to 105 trillion digits on a system supported by nearly a petabyte of Solidigm QLC SSDs. This achievement set a new benchmark for scale and demonstrated how efficiently a single on-premises machine could operate. A few months later, we did it again, this time reaching 202 trillion digits. This validated that high-density flash and careful tuning could outperform hyperscale infrastructure for this specific, demanding workload. Of course, records invite challengers. Linus Media Group and KIOXIA grabbed the crown with a 300 trillion-digit run, powered by a large Weka shared-storage cluster with 2PB of flash. That effort showed what storage-heavy traditional infrastructure could do, albeit with a rack of hardware, a large power bill, and cooling complexities. We couldn’t idly stand by and let that record stand! StorageReview has now pushed π to 314 trillion digits, TB Micron 6550 Ion SSDs. With the system build and tuning taking place in July, we kicked off our run on July 31, 2025. As luck would have it, our pi run wrapped during our second day of SC25, fitting for a new HPC record.Scaling y-cruncher to 314 Trillion Digits Once you cross into the hundreds of trillions of digits, y-cruncher behaves less like a traditional benchmark and more like a long-haul infrastructure stress test. The application itself is straightforward, but the way it interacts with hardware at this scale becomes the determining factor. Everything comes down to how well the system can keep thousands of multi-precision operations moving without stalling the CPUs or overwhelming the storage layer. The storage layer, specifically, is where this record was actually won. We deployed 40 Micron 6550 Ion Gen5 NVMe SSDs, 34 of which were allocated to y-cruncher. This SSD pool delivered about 2.1 petabytes, giving y-cruncher enough space to perform the 314T run. The remaining 6 SSDs were used to build a software RAID10 volume, which we used to record the 314T digits of pi. Design changes between the 16th- and 17th-generation Dell PowerEdge servers also contributed to improving the performance of our most recent 314T record run. In our previous 202T record, we leveraged the 24-bay Dell PowerEdge R760, which used a PCIe switch on the drive backplane, trading drive density for performance. On the 17th-generation Dell PowerEdge servers, such as the Intel-based R770 or AMD-based R7725, the backplanes reverted to direct-connect-only, with 2 or 4 PCIe lanes per bay. In the PowerEdge R7725 with its 40-bay Gen5 E3.S backplane, each SSD gets 2 PCIe lanes. While that may seem like a performance hit, the platform is still capable of reaching up to 280GB/s read and write when all 40 bays are hit simultaneously. Using the internal y-cruncher storage benchmark, we recorded the storage performance of each platform in the configuration used for each run. Across each workload, we saw storage performance increases ranging from 72% to 383%, with balanced read and write metrics. | 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 isn’t just a storage box; it also offers significant compute potential as a dual-socket AMD Turin platform. We leveraged 192-core AMD EPYC 9965 CPUs, totaling 384 cores. We also replaced the stock air-cooled heatsinks with liquid-cooled CoolIT SP5 cold plates, which were cooled by a CoolIT AHx10 Liquid-to-Air CDU. This combination kept the CPUs in higher sustained clock speeds, chassis fans operating at around 30% PWM, and our average system power consumption hovering around 1,600W. On the software side, the platform ran Ubuntu 24.04.2 LTS Server instead of Windows Server as in past runs. This maximized system stability and delivered significant gains in workload performance. We performed numerous test iterations before committing to kicking off our run, including reserving 4 out of 384 cores for background system operations. The result wasn’t just beating the existing pi record; we obliterated it across numerous metrics. Nothing comes close to our run in terms of performance, power consumption, and most impressively, reliability. We are also the only large-scale pi world-record run witho