MLCommons has published MLPerf Storage v3.0, adding checkpointing and inference workloads. The results are a useful reminder that AI storage is bought on the pipeline, not on peak read bandwidth.
What was announced
- MLCommons published MLPerf Storage v3.0 on 1 September 2026, with 19 organisations submitting 143 results, 11 of them first-time submitters.
- v3.0 covers the full pipeline for the first time: training throughput, checkpointing, and two new inference workloads covering vector database and KV cache.
- Everpure (formerly Pure Storage) posted the round's largest checkpoint figure with FlashBlade//EXA at 877.5 GiB/s write and 833.0 GiB/s read at 30 data nodes.
- Azure Managed Lustre was the first hyperscale cloud storage service submitted to the benchmark, posting 642.2 GiB/s checkpoint write from a 4,096 TiB deployment with 128 clients.
- YanRongTech's F9000X led 3D U-Net training at 543.9 GiB/s read feeding 99 simulated B200 accelerators; HPE was the only established enterprise array vendor to submit training results, with the K3000 at 345.8 GiB/s.
- v3.0 moved its simulated accelerators from H100 to B200, so v3.0 results are not comparable with v2.0.
- DDN, Huawei, Hammerspace and Lightbits, which defined the v2.0 leaderboard, did not submit this round.
The ETON view
The important change here is not who won, it is what is being measured. Checkpointing is now a first-class workload because idle accelerators during a checkpoint flush are the most expensive thing in an AI cluster. If you are specifying storage for training, the question to ask a vendor is no longer sustained read bandwidth, it is how long the cluster stalls while state is written, and how that scales as you add nodes.
Two caveats worth carrying into a procurement conversation. These are audited vendor submissions on tuned configurations, and half the leaderboard is made up of names that will not appear on a UK support contract, so a headline figure from a thirty-node cluster tells you very little about the four-node system you are actually costing. And the absence of several previous leaders means the table is not a market ranking. Use the benchmark for the shape of the problem, checkpoint write, vector and KV cache behaviour, then size real hardware against your own model and node count. For most of the AI storage builds we quote, the win comes from NVMe tiering and network design rather than from buying the array at the top of a chart.
Related infrastructure
- Hard Drive Options
- Disk Storage & Enclosures
- HPE Alletra Storage MP
- GIGABYTE Storage-Optimised Rack Servers
Category: Storage · Vendor: Everpure, Microsoft Azure, HPE, NVIDIA · Technology: MLPerf Storage v3.0, checkpointing, Lustre, NVMe, KV cache · Last verified: 02 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
