The committee cited architectural disclosure and file-system availability; its reported statement did not mention sanctions
The IO500 Committee has moved two SCNet submissions built on Sugon's ParaStor F9000 all-flash storage from the ISC26 Production rankings to the Research rankings, restoring Argonne National Laboratory's Aurora DAOS results to No. 1 on both the Production and 10 Node Production lists.
The committee has not published its explanation on io500.org. In a statement quoted by Tom's Hardware on September 11, it said the Sugon submission did not meet the highest reproducibility level because of a "lack of widely available architectural details and limited general availability of the file system."
The statement refers to the "Sugon ISC26 submission" in the singular, although two SCNet entries changed lists: the 500-client-node AICS-A and the 10-client-node AICS-B. The committee's news page contains no announcement after its May 2026 call for ISC26 submissions, and it has not published the reclassification date.
The benchmark results themselves remain listed as verified IO500 submissions. As of September 13, AICS-A ranks fourth on the Research list at 79,110.10, while AICS-B ranks fourth on the 10 Node Research list at 7,839.30.
Within those lists, AICS-A scores about 1.83 times Aurora's Research result of 43,218.80, and AICS-B scores about 2.09 times Aurora's 10-node Research result of 3,748.85.
Those are the relevant comparisons under IO500's messaging policy, which requires comparisons to use a single metric on a specific list and prohibits comparisons between Research and Production. The 2.46 figure Tom's Hardware reported instead divides AICS-A's Research score by Aurora's Production score of 32,165.90. The arithmetic is correct, but the comparison crosses lists with different eligibility requirements.
The reclassification did not erase the SCNet scores or invalidate the benchmark runs. It changed the category in which the results may be cited and the claims that can be attached to them.
IO500 says a result appearing on a published list must be described as "verified." According to its messaging policy, that means the committee audited the submission package for correctness, feasibility, and consistency. Production placement requires more: a system must meet the highest reproducibility level, operate as a production system, and have no single point of failure.
The distinction dates to the committee's April 2022 List Split Proposal. The committee said a single ranking placed systems with very different availability and durability characteristics side by side, leaving practitioners unable to tell whether a result came from a resilient production deployment or a configuration optimized for research and benchmarking.
Its companion Reproducibility Proposal established four levels: Undefined, Limited, Proprietary, and Fully Reproducible. A Proprietary submission supplies the required metadata and questionnaire but uses a system that is neither open source nor commercially available. A Fully Reproducible submission must also describe a system widely available without provider-imposed restrictions.
Open source is not mandatory. The proposal explicitly allows software distributed through a commercial license, and says the committee does not intend to require access to a submitter's storage system. It does, however, require the scripts, code, configuration files, and documentation needed to recreate the benchmark environment on appropriate hardware and software.
The current submission rules make Fully Reproducible entries eligible for Production if they also satisfy the other Production requirements. The committee's list definitions describe a Production System as one that runs real scientific, industrial, or business applications regularly and continuously, usually over a period measured in years, while tolerating any single component failure without manual intervention.
The highest reproducibility level is therefore necessary for Production placement but not sufficient on its own. HPE's ISC26 HPE-K3000-NDR DAOS result, for example, provides public reproducibility material but uses an unprotected configuration expressly submitted to the Research list. It currently ranks No. 22 there.
Commercial licensing is also not a bar to Production placement. JD Explore Academy's reproducibility questionnaire describes JPFS as commercially available, and its entry ranks No. 11 on the main Production list and No. 4 on the 10 Node Production list.
Aurora's Production questionnaire identifies DAOS version 2.4.0-2 and links to the software's source repository, packages, and documentation. It also points to a public reproducibility repository containing the server YAML and IO500 configuration used for the run.
That result used 642 of Aurora's 1,024 DAOS storage servers and 300 compute nodes, each with eight HPE Slingshot network interfaces. Bulk data used either erasure coding or two-way replication, while metadata was replicated. The questionnaire says the configuration could tolerate the loss of one server.
DAOS has operated under the DAOS Foundation within the Linux Foundation since November 2023. Its public repository uses the BSD-2-Clause Plus Patent License.
SCNet's AICS-A questionnaire describes ParaStor as a commercially available parallel storage system that research institutions and enterprises can acquire. It identifies ParaStor F9000 and links to a Sugon product-category page.
The questionnaire provides several architectural details. Each storage node has 12 NVMe SSDs and four 400-gigabit network ports. Data uses 14+2 erasure coding, while metadata is triply replicated. It also describes a kernel-mounted client and says the system can tolerate two simultaneous node or disk failures.
For the benchmark environment, however, the questionnaire says tuning artifacts and configuration files "can be shared upon request or included in the submission package." Unlike Aurora's submission page, the public AICS-A page has no Files tab.
The public record does not show whether the committee's review turned on that artifact language, the level of detail on the product page, information supplied privately, or a combination of those factors. The reported statement identifies only the broader problems of architectural disclosure and general availability.
Storage architect Glenn K. Lockwood, who has written extensively about IO500, highlighted the list change on social media on September 10. In his July ISC26 recap, he had described ParaStor as "a real product" and noted that Sugon displayed three F9000 racks at the conference. He also said Sugon had not disclosed how the system handled metadata and called for independent testing of several claimed capabilities.
A conference display supports the claim that hardware exists, but it does not by itself establish general commercial availability or provide the architectural and configuration material required for IO500's highest reproducibility level.
SCNet and Sugon play different roles in the submission, but they are not unrelated. SCNet is the National Supercomputing Internet platform and the institution named by IO500. Sugon is the storage vendor and developer of ParaStor.
China launched the national platform in Tianjin in April 2024 after beginning its construction a year earlier, according to a Chinese government report citing the Ministry of Science and Technology. An SCNet platform agreement identifies Sugon Intelligent Computing Information Technology as its operator, and an SCNet vendor page describes that company as a wholly owned Sugon subsidiary.
That relationship explains why IO500's reported statement refers to Sugon even though the list names SCNet as the submitting institution. It does not resolve why the statement uses the singular when two entries were reclassified.
Pengcheng Laboratory's CloudBrain illustrates the Research rules more directly. Its questionnaire describes the storage software as an in-house implementation whose source is not publicly available. It also says it disabled high-availability and data-redundancy features for the benchmark.
The entry carries the Proprietary reproducibility designation and leads the Research list at 603,334.56 on Huawei OceanStor A800 hardware. That is about 7.6 times AICS-A's overall score despite delivering less than one-third of AICS-A's bandwidth: 8,291 GiB/s against 26,888 GiB/s. CloudBrain's metadata score of 43.9 million kIOPS drives the difference.
The two 300-client-node Aurora entries compared here ran at SC23 against the same 642 DAOS storage servers, but used different protection schemes and process counts.
Aurora's Research configuration used no data protection and 31,200 client processes. Its questionnaire says losing one server would make data unavailable. It scored 43,218.80.
The Production configuration used protected bulk data and replicated metadata with 62,400 client processes. It scored 32,165.90. Because both the protection settings and process counts changed, the two pages do not isolate the performance cost of data protection.
Aurora also has separate 10-client-node Research and Production results.
The runs occurred while Aurora was in a preproduction environment. IO500's definition permits a machine that "is or will be" used for continuing production applications, allowing systems approaching general service to qualify if they otherwise meet the requirements.
As of September 13, the IO500 site displays AICS-A and AICS-B on the Research lists with Fully Reproducible badges. Pengcheng's CloudBrain entries display Proprietary badges.
That creates an unresolved inconsistency. The committee's reported statement says the SCNet submission did not meet the highest reproducibility level, yet the site still shows the badge associated with that level.
The field may be stale, or the committee may have left the assigned badge in place while separately deciding that the entries were not eligible for Production. IO500 has not publicly explained the discrepancy. A Fully Reproducible badge does not guarantee Production placement, as the HPE research entry shows, but the badge and the committee's stated reason still need to be reconciled.
There is also a smaller numerical discrepancy. The ISC26 Full list displays AICS-A at 79,110.05, while the Research list and submission page show 79,110.10. The committee statement quoted by Tom's Hardware uses 79,110.05.
Sugon has been on the US Commerce Department's Entity List since June 24, 2019. The relevant Federal Register notice imposed a license requirement for all items subject to the Export Administration Regulations and a presumption-of-denial review policy.
The notice identified Sugon, the Wuxi Jiangnan Institute of Computing Technology, and the National University of Defense Technology as leaders in China's exascale HPC development. It also said Sugon had acknowledged military end uses and users of its high-performance computers.
Nothing in the committee statement quoted by Tom's Hardware mentions the Entity List, sanctions, or export controls. That establishes what the committee publicly cited, not every factor that may have informed its review.
IO500's record shows that the US listing has not automatically excluded Sugon-related submissions. Its news archive records the Sugon Cloud Storage Laboratory winning the 10 Node Challenge at ISC22 and SC22, three years after Sugon joined the Entity List. The SC22 ParaStor entry remains on the 10 Node Research list.
SCN therefore has no evidence that sanctions drove the current reclassification. The supportable conclusion is narrower: the committee's reported explanation cites reproducibility, architectural disclosure, and general availability.
Sugon's June 24 press release said ParaStor F9000 had achieved No. 1 on both Production lists and that the result demonstrated leading performance under production conditions. The release calls the event "SC26" rather than ISC26. Its ranking claim accurately described the lists at the time but no longer matches their current composition.
The IO500 overall score is the geometric mean of a bandwidth score and a metadata score, each of which is itself a geometric mean of several tests. As Lockwood explains, this gives the bandwidth component, measured in GiB/s, the same weight as the metadata component, measured in kIOPS.
That construction can make the overall score difficult to interpret without its components. AICS-A exceeds Aurora's Research configuration in both: about 2.37 times its bandwidth and 1.42 times its metadata result. CloudBrain, by contrast, obtains its much higher overall score from extraordinary metadata throughput despite substantially lower bandwidth than AICS-A.
The SCNet results remain significant published measurements. Their move to Research does not invalidate the measured figures. It means IO500 does not currently consider the submissions eligible for the assurances attached to Production placement: general availability, the highest level of reproducibility, production use and resilience against a single component failure.
IO500's lists page also describes a route back. Submitters seeking to move an entry from Research to Production may contact the steering committee. Neither SCNet nor Sugon has publicly said whether it plans to provide additional material or request another review.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
