Joint engineering between Samsung and Hammerspace resulted in record IO500 benchmark gains, including 2X higher bandwidth and 34% overall performance increase, reflecting the ongoing performance advancements in the Hammerspace Data Platform
Summary:
- These results demonstrate – for the first time in the IO500 10-Node Production list – true HPC-class performance using only standard hardware and software. No proprietary clients, drivers or kernel modules, just standard Linux and NFS on enterprise-grade server, storage, and network equipment
- Extreme performance was achieved using open, widely understood Linux/NFS architectures, eliminating the specialized expertise historically required for HPC filesystems
- The overall score for Samsung’s entry with Hammerspace in the 10-Node Production category was 33.7% higher than the previous 10-Node Research results submitted by Hammerspace just five months ago at ISC ’25.
- The higher overall score was largely driven by improvements in all bandwidth test categories, with the overall bandwidth score increasing by 100.4%. Notably, performance on the “IOR-Hard-Write” and “IOR-Hard-Read” tests increased by a solid 21.1% and stunning 809.5% respectively
Samsung partnered with Hammerspace to run the 10-Node IO500 benchmark on Samsung Memory Division’s production AE-25 system. Samsung researchers use the AE-25 to run simulation and benchmark tasks essential for next-generation memory product development, including SSDs, DRAM, and CXL. The Hammerspace Data Platform provided high-performance, standards-based shared storage to the cluster for the benchmark runs, backed by Samsung PM1753 Enterprise NVMe SSDs.
“Our work with Hammerspace demonstrates the performance headroom that still exists in modern storage when software and hardware innovation advance together,” said Hojun Shim, Vice President of Memory Application Engineering at Samsung Electronics. “By combining Samsung’s enterprise-class NVMe SSDs with the Hammerspace Data Platform, we’re setting new standards for how data is managed and accelerated in AI and HPC environments.”
The IO500 Benchmark
The IO500 is an open benchmark suite designed to evaluate how storage systems handle common scientific and HPC workloads. It is now also being looked at closely by hyperscalers, neoclouds, and enterprises as they build out their AI infrastructure.
The tests measure systems based on both bandwidth and metadata performance to provide a holistic view. The detailed test results are summarized in three numbers; a composite bandwidth figure, a composite IOPS figure, and a calculated overall numeric score.
Benchmark Results and Discussion
| Samsung/Hammerspace results, 10-Node Production Cluster, SC25 | ||
| Overall Score | Bandwidth GiB/s | IOPS kIOPS |
| 85.23 | 74.66 | 97.29 |
This score puts Hammerspace at number 18 in the 10-Node Production ranking, alongside conventional HPC file system architectures like Lustre, DAOS, and WEKA – but without the requirement for specialized networking, proprietary clients, and specialized hardware.
Producing results like these historically required a proprietary HPC file system. Traditional HPC file systems are labor-intensive to configure and maintain, require custom software on every client that needs access, and lack features enterprises require. There is a better way.
Hammerspace and Samsung delivered the impressive numbers above using 10 industry-standard Linux clients with conventional IP-over-IB networking. Storage was provided by Samsung PM1753 NVMe SSDs housed in Hammerspace DSX nodes. Client machines connected to the storage system using the standard Linux NFS client — no custom client software or kernel patches were required.
This is the future of AI and HPC infrastructure. Scalable high performance from simple-to-deploy, standards-based systems with no vendor lock-in.
To successfully launch AI and HPC projects, enterprises need low-friction infrastructure based on familiar hardware and software they already own – not complex proprietary systems that require a roomful of PhDs to maintain.
Using a simple, standards-based approach makes it easy to leverage cutting-edge storage technology such as the Samsung PM1753 SSDs used in this benchmark. The PM1753 is a low-power TLC SSD that uses eighth-generation V-NAND technology. It was expressly designed to meet the high throughput demands and heavy write loads of AI applications while minimizing power consumption.
Current Results (SC25) vs. Previous Results (ISC25)
The table below compares the overall results from Hammerspace’s previous IO500 submission at ISC25 with the results from this run with Samsung. Even though there were significant differences in hardware, software, and configuration between the two setups, some interesting and meaningful comparisons may still be drawn.
| Measure | Previous Score (ISC25) | Current Score (SC25) | DifferenceISC25 🡪 SC25 |
| Overall Total | 63.77 | 85.23 | +33.7% |
| Overall Bandwidth GiB/s | 37.25 | 74.66 | +100.4% |
| Overall Metadata kIOPS | 109.16 | 97.29 | -10.9% |
The overall score increased by over one-third, a huge achievement. This increase was due to improved bandwidth scores across the board, as well as an improved Find kIOPS score.
Within the improved bandwidth scores, the standout was the IOR-HARD-READ score, which increased by over 800% versus previous, while the IOR-HARD-WRITE score increased 21.1%. IOR-EASY-WRITE and IOR-EASY-READ also bumped up 21% each.
Recent Linux kernel enhancements engineered and contributed upstream by Hammerspace were responsible for the improved scores, as were major improvements in Hammerspace code. Configuration differences between the test setups also certainly made a difference.
Overall, metadata (kIOPS) scores were slightly down compared to the 10-Node Research submission at ISC25. This was due to the use of redundant Anvils in the Production configuration at Samsung.
The Research challenge submission was designed for a scratch use case which valued performance over protection. In contrast, the Production system at Samsung used a second metadata server to provide the redundancy and protection that the vast majority of real-world deployments require. This additional layer of resiliency introduces a small amount of latency, which naturally impacts metadata performance.
In other words, the Research configuration was optimized purely for maximal metadata performance, while the Production system at the Samsung Memory Research Center is a hardened architecture where durability and robustness take precedence over absolute peak metadata speed.
IO500 Test Configuration
The test configuration was typical for a Hammerspace high-performance parallel file system deployment. Two Hammerspace Anvil nodes in a high availability configuration, 10 Hammerspace DSX nodes housing the Samsung SSDs, and 10 Linux clients running the benchmark software. All servers are powerful but standard systems from Dell and Supermicro, connected with 400Gb InfiniBand through an NVIDIA switch.
With Hammerspace, Anvil nodes are responsible for metadata. DSX nodes may serve multiple roles, including as hosts for block storage, portals for file and object protocols, and policy-controlled data movers. For this benchmark run, the DSX nodes were configured to operate solely as hosts for the Samsung SSDs. Hammerspace unified the storage across all DSX nodes into a single shared filesystem, which clients mounted with parallel NFS (pNFS) v4.2 using flexible files layouts.
The pNFS v4.2 protocol separates the metadata path from the data path, similar to legacy HPC file systems. This architecture is key to achieving high performance. Clients communicate with a metadata server to request authorization to perform I/O, and in return receive a “layout.” The information in the layout enables the client to go directly to the storage servers to read and write data. There is no need to proxy all I/O through a “NAS head,” secondary network, or other bottleneck.
Hammerspace Data Platform
Keep in mind that Hammerspace is a comprehensive data platform, not just a high-performance file system. It can unify an organization’s entire data estate across silos, sites, and clouds, managing data placement, protection, and governance with a single set of policies. Data orchestration driven by customizable metadata powers intelligent, automated workflows.
Samsung Memory Research Center (SMRC)
All systems deployed for the Samsung IO500 benchmark with Hammerspace were built on the latest servers and networking equipment supplied by the Samsung Memory Research Center (SMRC). The SMRC is an open, collaborative platform that brings together customers and partners to evaluate Samsung’s latest memory technologies in real-world datacenter environments. By providing seamless technical support, a validated infrastructure, and future-ready systems, SMRC accelerates technology adoption and maximizes the value of cutting-edge memory solutions.
More About the IO500
Samsung’s entry with Hammerspace was in the 10-Node Production category. IO500 participants range from small organizations to huge government labs, with entries in four categories.
| Category | Who Can Submit | Client Count | Storage Node Count |
| Production | Production Customer | Any | Any |
| 10-Node Production | Production Customer | 10 | Any |
| Research | Anyone | Any | Any |
| 10-Node Research | Anyone | 10 | Any |
There are no bounds on client size, storage system size, or cost, which leads to a wide range of scores. The ”10-Node” categories introduce a limit of 10 on the client count (the machines that run the benchmark code) as a way to compare setups that are a little more constrained. It is noteworthy that the storage system in Samsung’s submission also contained only 10 nodes, delivering excellent performance in a very space, power, and cost-efficient package.
The distinction between Production and Research is important. Production systems “have a day job” doing work in academia, industry, and government. These configurations highlight what works in the real world. Research systems, in contrast, may showcase experimental technology or approaches that are not hardened, and may even be constructed solely to do well on the IO500 benchmark.
The benchmark consists of a range of tests focusing on different workload patterns summarized in the table below. Tests involve read, write, delete, stat, and find operations. Results are output for each test and then combined and summarized.
| Workload | Details | AI Workload Relevance |
| IOEasy | File per process, large block sequential IO | Training data loading, model saving |
| IOHard | Shared single file for all processes, 47KB+ IO | Small, random unaligned IO typical in optimizer state and KV cache updates |
| MDEasy | Directory per process, zero-length files | Pure metadata speed test |
| MDHard | Small (3901 byte) files, single shared directory for all processes | Mimics checkpoint/restore logic, where file create/delete dominates |
| Find | Finding relevant objects based on patterns | Metadata search operation – custom Find algorithms are allowed |
Conclusion
Samsung’s IO500 entry with Hammerspace demonstrates how a truly standards-based, Linux-native file system can now deliver the extreme performance required for today’s largest HPC and AI compute environments. Hammerspace delivers the performance these environments demand while staying aligned with open standards and upstream Linux innovation.
But the larger significance of this benchmark goes far beyond a single set of results.
It reinforces that the community’s long-term bet on Linux, NFSv4.2, and pNFS is paying off — and that with the right architecture, standard protocols can achieve the low latency, massive parallelism, and global scalability required for AI training, RAG pipelines, agentic workflows, and classical HPC simulations.
Hammerspace is uniquely engineered to take full advantage of these upstream kernel and protocol advancements, translating them into real, measurable performance at scale. As HPC and AI continue to converge on shared compute infrastructure, this benchmark validates a new reality: open, standards-based data architectures are now fully capable of powering the next generation of performance-intensive computing.
Learn More
- Read the press release: “Hammerspace Breaks IO500 Barriers: First Standards-Based Linux + NFS System To Achieve True HPC-Class Performance”
- Listen to the Data Unchained Podcast: “Hammerspace Breaks IO500 Barriers: How They Built the Fastest NFS-Based Benchmark Ever w/ Jon Flynn“
- Read the full details of the Samsung / Hammerspace IO500 submission at the IO500 website
- Contact Hammerspace’s team of experts to book a meeting, set up a 1:1 demo or take a self-guided tour.
