Sugon Tops IO500, Showcases AI Infrastructure in Europe

The IO500 production list, which restricts entries to systems deployed under real operational conditions with redundancy requirements, is a more operationally meaningful benchmark than the research category. For AI infrastructure teams: a 79,110 IO500 score with 26,888 GiB/s bandwidth and 232,754 kIOP/s metadata throughput, confirmed by io500.org - the independent benchmark organization - points to a system capable of sustaining high parallel I/O across large GPU clusters. The 2.4x score gap over the second-ranked entry (Argonne National Lab's Aurora, using Intel DAOS at 32,165) is substantial. Procurement teams should still seek published run configuration details and independent customer references before drawing architecture conclusions.

Per PR Newswire and StorageNewsletter, Sugon showcased its AI and advanced computing portfolio at ISC High Performance 2026 in Hamburg on June 24, 2026. The presentation covered AI supernodes, AI superclusters, the ParaStor F9000 high-performance storage system, RDMA networking, liquid-cooled data center solutions, servers, workstations, and computing services.

Per PR Newswire, Sugon's ParaStor F9000 distributed all-flash storage system achieved No. 1 on both the IO500 Production SC26 List and the 10 Node Production SC26 List. io500.org, the independent benchmark organization that maintains the IO500 list, lists the ISC26 production top entry under institution SCNet as the submitter, with Sugon as the storage vendor and ParaStor as the file system. The recorded score is 79,110.05, with 26,888.39 GiB/s bandwidth and 232,754.76 kIOP/s metadata operations per second, using 500 client nodes and 64,000 total client processes. The second-ranked production entry, Argonne National Laboratory's Aurora (Intel DAOS), scored 32,165.90 - making the Sugon-powered system's score 2.46x higher. Per PR Newswire and StorageNewsletter, the result builds on Sugon's prior IO500 track record, including a previous No. 1 finish in the 10-Node Challenge at SC22.

IO500 production rankings emphasize sustained real-world throughput and metadata performance under operational constraints such as redundancy and workload mix. Teams running large training clusters commonly see aggregate throughput limited by storage latency and parallel I/O scalability rather than raw GPU compute. Vendors reporting production IO500 leadership typically pair software-defined distributed file systems with NVMe-oF or RDMA networking and all-flash hardware to improve both bandwidth and I/O consistency across nodes. The participation of a major national lab (Argonne) as second-ranked provides useful context: the gap between first and second here is atypically large by IO500 historical norms.

Implications for practitioners

• •For training cluster architects: a production-class all-flash distributed storage system with an IO500 lead of this magnitude can reduce variance in epoch times and improve effective GPU utilization when training on sharded datasets at scale. This is a general industry implication; the SCNet submission context means practitioners should verify how the system maps to their specific deployment architecture.
• •For infra procurement teams: liquid cooling and RDMA networking lower operational barriers for dense GPU racks. Seek published run artifacts (configuration details, SSD models, network topology) and independent customer case studies before finalizing procurement decisions.

Observers should look for: published IO500 run details and questionnaire data for the SCNet submission (available at io500.org/questionnaires/view/803); independent customer case studies involving ParaStor F9000 at similar scales; and whether the claimed 10 Node Production SC26 List top ranking also appears in the official io500.org 10-node production tab. Per PR Newswire, Sugon was founded in 1996 and develops computing infrastructure for enterprises, research institutions, and public-sector organizations.