Western Digital validates AI-optimized storage with MLPerf V2 results

To validate real-world performance, Western Digital partnered with SDS provider PEAK:AIO and used high-performance KIOXIA CM7-V Series NVMe SSDs in the OpenFlex Data24 system, enabling sustained data delivery across multiple simulated GPU nodes

Western Digital has announced its MLPerf Storage V2 benchmark results, highlighting the high-performance capabilities of its OpenFlex Data24 4000 Series NVMe-oF Storage Platform for demanding AI workloads. The results validate the solution’s efficiency, scalability, and cost-effectiveness in supporting modern AI infrastructure.

As AI applications grow more complex and compute-intensive, ensuring that storage systems can keep pace has become critical. Western Digital’s OpenFlex Data24 platform extends NVMe flash performance over Ethernet fabrics, delivering low-latency shared storage designed for disaggregated and scalable AI environments. By allowing compute and storage to scale independently, it enables more flexibility and efficiency in AI deployments.

To ensure the solution meets real-world demands, Western Digital collaborated with software-defined storage (SDS) provider PEAK:AIO and used KIOXIA CM7-V Series NVMe SSDs for the benchmark. These SSDs, known for high performance under heavy AI workloads, enabled sustained data delivery to multiple simulated GPU nodes when deployed in the OpenFlex Data24 system.

Benchmarking next-gen AI storage

The MLPerf Storage V2 benchmark, a recognized industry standard, evaluates storage system performance under AI training workloads using simulated GPU client nodes. Two key AI benchmarks—3D U-Net and ResNet-50—were used to measure throughput, latency, and concurrent access performance.

In the 3D U-Net test, which stresses storage systems with large 3D datasets, OpenFlex Data24 achieved a sustained read throughput of 106.5 GB/s across 36 simulated NVIDIA H100 GPUs. When integrated with PEAK:AIO’s AI Data Server, it still delivered 64.9 GB/s, saturating 22 GPUs—demonstrating its strength in bandwidth-intensive environments.

For the ResNet-50 benchmark, designed to assess balanced compute and data movement, OpenFlex Data24 performed optimally across 186 simulated H100 GPUs, showing excellent drive utilization and efficiency. The PEAK:AIO integration again showed strong results, supporting 52 GPUs from a single head server and client node.

“These results affirm OpenFlex Data24 as a cornerstone for next-gen AI infrastructure,” said Kurt Chan, VP and GM, Western Digital Platforms. “It maximizes GPU performance while reducing cost and complexity.”

PEAK:AIO CEO Roger Cummings added, “Together, we’re proving that high-performance AI infrastructure can be efficient, scalable, and simple to deploy.”

Western Digital’s OpenFlex Data24, with RapidFlex adapters supporting up to 12 host connections without a switch, offers a streamlined path for organizations scaling AI workloads across various deployment sizes.