News Summary
- Huawei OceanStor A series storage once again ranked No. 1 worldwide in the MLPerf Storage v2.0 benchmarks, leading in performance per system, per rack unit, and per client.
- In the 3D U-Net training workload, OceanStor A series achieved a global record of 698 GiB/s, sustaining GPU utilization above 90%.
- In the checkpointing benchmark, OceanStor delivered 6.7x higher performance than the second-best performer, reinforcing its leadership in large AI model training.
OceanStor A Series Secures Global Leadership in AI Storage
On September 15, 2025, Huawei announced that its OceanStor A series storage retained the top position in the MLPerf Storage v2.0 benchmarks, the industry’s authoritative standard for AI storage performance.
Conducted by MLCommons in collaboration with the Jinan Institute of Supercomputing Technology, the results highlight OceanStor’s dominance in supporting the growing computing demands of large-scale AI clusters.
The OceanStor A series ranked first globally across multiple categories, including performance per system, per rack unit, and per client. In the 3D U-Net workload - a test that measures bandwidth and GPU utilization for model training, an 8U dual-node OceanStor A800 system delivered 698 GiB/s of stable bandwidth across 255 NVIDIA H100 GPUs, while the A600 system achieved 108 GiB/s per rack unit and 104 GiB/s per client.
In the newly introduced checkpointing test, which evaluates resumable training and model archiving scenarios, OceanStor A series delivered record-breaking results. For example, with the Llama3-70B workload, it reached 68.8 GiB/s read and 62.4 GiB/s write bandwidth—6.7 times higher than the next-best competitor.
Innovations Driving AI Performance
Huawei OceanStor A series is purpose-built to accelerate the adoption of large AI models and high-performance computing (HPC). Key innovations include:
- Massive Scalability: EB-level capacity with hundreds of TBs of cluster bandwidth.
- High Reliability: 99.999% availability through architectural innovation.
- Performance Optimizations: PB-level KV cache pool reduces Time to First Token (TTFT) by up to 90%, boosting inference throughput by more than 10x in long-sequence scenarios.
- AI-Ready Design: Built-in Retrieval-Augmented Generation (RAG) knowledge base with multi-mode retrieval (scalars, vectors, tensors, graphs).
These advancements ensure OceanStor A series can handle the exponential growth in data, providing end-to-end acceleration for training and inference.
Availability
The Huawei OceanStor A series is available now, with continuous innovations planned to further support HPC and large AI model workloads. Future updates will continue to optimize training performance and broaden AI deployment capabilities.
Conclusion and Network-switch.com’s Perspective
Huawei’s achievement in the MLPerf Storage v2.0 benchmarks represents a significant milestone for enterprise AI infrastructure. By consistently leading in storage performance, OceanStor A series provides the reliability and scalability needed to support next-generation AI workloads.
This leadership not only accelerates the adoption of large models but also demonstrates the vital role of storage in ensuring efficient, resilient, and scalable AI training environments.
At Network-switch.com, we view Huawei’s accomplishment as a clear testament to its innovation and engineering strength. OceanStor A series sets a global benchmark for AI-ready storage, and its continuous evolution inspires confidence in enterprises pursuing AI-driven transformation.
As a professional third-party distributor, we are proud to collaborate with leading partners like Huawei, and we remain committed to delivering world-class solutions that empower our customers to thrive in the intelligent era.