AI BENCHMARK MLPERF Sep 02, 2026
MLCommons published the v3.0 MLPerf Storage results round on Sep 1 (Blocks & Files coverage), the first big restructure of the benchmark since it launched around training workloads (Unet3D, Cosmoflow, Resnet50 on A100/H100-class GPUs). v3.0 defines four workloads: training (Unet3D/RetinaNet pushing for maximum accelerators above 90% utilization, simulated with Nvidia B200 and AMD MI355 GPUs); checkpointing with Llama-3-class models from 8B to 1.25T parameters, timed as 10 write passes followed by 10 read passes; a VDB vector-database test (Milvus, 1M vectors at 1,536 dimensions, maximizing queries/s while reporting latency and recall); and a KV-cache test measuring how many concurrent inference conversations the storage layer can sustain. It also adds an S3 object-storage data layer alongside POSIX — roughly one-sixth of submissions in this round used S3.
Nineteen suppliers ran tests. The notable absence list: DDN, Dell, Huawei, IBM, NetApp, VAST Data and WEKA. MLCommons itself flags that "traditional NAS is thin" and the legacy enterprise incumbents are mostly missing, and only four suppliers (NewFW, Samsung, Suzhou Zishan, TTA) submitted VDB results. Everpure's FlashBlade//EXA ranked first across the 405B and 1.25T checkpointing and KV-cache categories, hitting 877.52 GiB/s write (17.74s) and 588.28 GiB/s read (28.99s) for the 1.25T checkpoint over 30 data nodes with 1,024 simulated accelerators. There is deliberately no single winner: the results spread across a ~145-row by 55-column spreadsheet (12 columns cover KV-cache alone), and MLCommons suggests normalizing by rack unit and by watt — on-prem submissions hit a median 14 GB/s/watt on checkpoint writes and 34 GB/s/watt on Unet3D reads.
The admin lesson: this round is the first public yardstick for inference-era AI storage — checkpointing at your model-size class, KV-cache, vector-DB — and raw IOPS/latency numbers no longer answer the question "can it checkpoint a 405B model fast enough to keep GPUs busy?" When a vendor quotes an AI array, ask which MLPerf Storage v3.0 workloads it submitted and at what parameter scale, then normalize per rack unit and per watt. NetApp has no v3.0 submission, which proves nothing about ONTAP performance by itself but does mean AFF/ASA AI claims currently rest on vendor-published tests — worth validating in-house with the same 10×-write/10×-read methodology at your real model size before signing. And the new S3 layer legitimizes object storage for checkpoint data (the working group expects object to become preferred as AI contexts scale into the trillions) — directly relevant if you are planning StorageGRID or ONTAP S3 tiers as the checkpoint target.