{"resourceId":"storagereview-mlperf-storage-v3-everpure-2026-09-01","versions":[{"version":"external-4f4b0f819aa7aeb06f83f2803e7c2b043b0ea5206e645b9493a4e42cbcd5ddfe","resource":{"id":"storagereview-mlperf-storage-v3-everpure-2026-09-01","title":"MLPerf Storage v3.0: 877 GiB/s Checkpoints, a Cloud First, and a Leaderboard Turned Over","organization":"StorageReview (reporting on MLCommons MLPerf Storage v3.0)","sector":"Independent storage trade press / audited industry benchmark","geography":"Global","publishedAt":"2026-09-01","publicationDate":"2026-09-01","eventDate":"2026-09-01","sourceName":"StorageReview","sourceLabel":"StorageReview news, September 1, 2026","sourceUrl":"https://www.storagereview.com/news/mlperf-storage-v3-0-877-gib-s-checkpoints-a-cloud-first-and-a-leaderboard-turned-over","evidenceClass":"independent-reporting","outcomeClass":"mixed","topics":["infrastructure","governance-procurement"],"finding":"In its first MLPerf Storage appearance, Everpure's FlashBlade//EXA posted the round's top checkpoint result (877.5 GiB/s write, 833.0 GiB/s read at 30 data nodes, scaling near-linearly from 327.6 GiB/s at 10 nodes) and top KV-cache read (1,623.4 GiB/s on 51 hosts), plus 85,736 tokens/sec (Llama 3.1 8B) and 33,403 tokens/sec (70B) storage-only. StorageReview stresses these are audited vendor submissions it did not verify, that v3.0 results are not comparable to v2.0, and that VAST Data, WEKA and DDN sat the round out, so the audited and deployment records point to different vendor lists.","sledRelevance":"For research universities and state HPC/AI centres, audited checkpoint and KV-cache results are better evaluation inputs than vendor multipliers, but the tested //EXA configuration (30 blades plus 30 data nodes, about 866 TB) is far larger than most SLED deployments.","evidence":"MLCommons-audited benchmark results relayed by an independent outlet; vendor-submitted configurations; no SLED deployment. Results show storage throughput under synthetic workloads, not end-to-end application outcomes.","architectureImplications":"The submitted design split metadata over pNFS/TCP and bulk data over NFSv3 on RDMA; buyers need RDMA-capable networking to approach these figures.","governanceImplications":"Supports evaluation criteria that ask for audited benchmark results with disclosed configuration, while noting absent competitors cannot be judged from the round.","securityPrivacyImplications":"No security content.","caveats":"Published September 1, 2026, before this stream's first run; included because it is the only audited evidence located that tests Everpure's AI storage claims and is directly relevant to the September 30 PureKVA claims. Competitor storage vendors are referenced only for benchmark context.","streamIds":["everpure"],"roles":{"sales":"Sales framing: this is the strongest third-party evidence available for Everpure AI storage, and it should be presented carefully. FlashBlade//EXA topped the audited checkpoint and KV-cache categories, but StorageReview stresses these are vendor submissions it did not verify and that several competitors did not submit. For SLED research buyers, offer to map the tested configuration to their proposed scale rather than implying the top-line numbers transfer to a smaller system.","engineering":"Engineering framing: read the configuration, not just the headline. The result used 30 FlashBlade//EXA blades with 30 Linux/NVMe data nodes, roughly 866 TB usable, pNFS/TCP metadata and NFSv3 over RDMA. Most campus or state clusters are far smaller and may lack RDMA fabrics. Use the near-linear scaling points (10, 15, 20, 30 nodes) to estimate expectations, and request results at the proposed scale. Remember v3.0 is not comparable with v2.0.","delivery":"Delivery framing: benchmark results do not become acceptance criteria automatically. Delivery should define site acceptance tests at the purchased scale, for example checkpoint write time for the customer's model sizes and KV-cache read throughput under expected concurrency, and confirm the network design matches what was benchmarked. Record differences between the audited configuration and the delivered one so stakeholders understand why measured results may differ."},"roleDetails":{"sales":{"customerProblem":"Research and state AI buyers struggle to compare AI storage vendors on credible, audited evidence.","stakeholders":["Research computing director","HPC architect","Procurement officer"],"discoveryQuestions":["What checkpoint sizes and frequencies do your training jobs use?","Do you have or plan RDMA networking?","What scale of storage are you actually buying?"],"valueHypothesis":"Audited results indicate FlashBlade//EXA can scale checkpoint throughput near-linearly; benefit at smaller scale must be estimated and tested.","potentialEngagement":"A sizing workshop mapping audited MLPerf configurations to the customer's planned scale and workloads.","unsupportedClaims":["That results transfer to small configurations","That absent competitors are slower","End-to-end application speedups"]},"engineering":{"fit":"Relevant for large training/checkpoint and KV-cache workloads at research-centre scale.","architectureIntegration":"FlashBlade//EXA with separate metadata (pNFS/TCP) and data nodes (NFSv3 over RDMA).","prerequisites":["RDMA-capable fabric","Linux NVMe data nodes","Workload profiles"],"constraints":["Tested scale far above typical SLED deployments","Vendor-submitted configurations","Not comparable with v2.0"],"security":"Not addressed by the benchmark; standard storage security review still required.","proofOfValue":["Checkpoint write time at purchased scale","KV-cache read throughput under expected concurrency","Network validation against benchmarked design"]},"delivery":{"work":"Define and run site acceptance tests that mirror benchmark categories at the delivered scale.","dependencies":["Network team","Representative models and checkpoints","Vendor configuration details"],"ownership":"HPC/research computing operations.","skillsAdoption":"Staff need benchmarking and RDMA networking skills.","governanceCheckpoints":["Acceptance test plan approval","Configuration deviation log","Results sign-off"],"acceptanceCriteria":["Proposed: checkpoint write throughput meets an agreed target at delivered scale","Proposed: KV-cache read throughput measured at expected concurrency","Proposed: deviations from audited configuration documented"],"risks":["Expectation gap from headline numbers","Network bottlenecks","Benchmark-to-workload mismatch"]}},"retrievedAt":"2026-10-30T03:08:11Z","enrichedAt":"2026-10-30T03:13:23Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"No accessibility or workforce findings.","procurementImplications":"Procurements can reference MLPerf Storage v3.0 categories (checkpoint, KV cache) and require submissions at a scale comparable to the proposed configuration.","operatingModelImplications":"Encourages separating audited benchmark evidence from deployment evidence in vendor evaluation.","peerReview":"unknown","sourceVerification":{"openedUrl":"https://www.storagereview.com/news/mlperf-storage-v3-0-877-gib-s-checkpoints-a-cloud-first-and-a-leaderboard-turned-over","referenceExcerpt":"These are MLCommons-audited vendor submissions; StorageReview did not conduct or independently verify this testing.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}