China's Pengcheng Cloud Brain III Smashes IO500 Record, Powered by Huawei Storage

There are benchmarks that increment, and then there are benchmarks that make you do a double-take. The Pengcheng Cloud Brain III — a supercomputing system built by Pengcheng Lab and powered by Huawei’s OceanStor A800 storage — just did the latter. It topped both the Full List and Research List of the IO500 benchmark, the global standard for storage performance in high-performance computing and AI workloads.

The numbers are hard to ignore. The system scored 603,334.58 points overall, with a bandwidth of 8,291.11 GiB/s and metadata performance of 43,903,983.64 kIOP/s. That’s a 2.8x improvement over the previous world record. Pengcheng Lab has now held the IO500 crown for twelve consecutive iterations.

The IO500 isn’t a simple read-and-write test. It runs a complex mixed-load model designed to push storage systems to their limits under extreme concurrency and large-cluster conditions. The Research List, in particular, requires coordinating hundreds of compute and storage nodes working in parallel to measure the theoretical peak performance of the hardware-software combination. Passing it means the system can sustain its speed under real-world pressure, not just in a controlled test environment.

For this run, Huawei deployed the OceanStor A800 across 664 compute nodes, firing 79,680 parallel processes through 13 test scenarios covering a wide range of workloads. The results validated the storage system’s performance and stability at an unusually large scale.

Huawei credits three design pillars for the A800’s performance.

The first is architecture. A peer-to-peer interconnect and shared network cards eliminate the bottlenecks that typically choke large-scale connections, pushing single-node shared bandwidth higher. A data-control separation design lets the DPU talk directly to the SSD, bypassing the CPU’s semantic translation overhead — achieving 500 GB/s throughput per enclosure.

The second is multi-dimensional scaling. A horizontal-and-vertical hybrid expansion model, built on a disk-controller integrated architecture, allows a single cluster to support up to 512 controllers with aggregate bandwidth exceeding 100 TB/s and EB-level storage capacity. The system also accommodates DPUs and NPUs for accelerated data processing and vector search offload.

The third layer is smart algorithms. Dynamic directory sharding and striping-parallel technology spread metadata and data loads across the entire cluster, eliminating single points of contention. A DataTurbo high-performance file acceleration engine, paired with intelligent multi-level cache eviction, tackles the performance degradation that typically plagues metadata operations on massive numbers of small files.

Huawei also noted that the Cheeloo-1 system — built on the OceanStor Pacific distributed storage — remains at the top of the 10-Node Research List.