From the Everpure edition of October 29, 2026
MLPerf Storage v3.0: 877 GiB/s Checkpoints, a Cloud First, and a Leaderboard Turned Over
StorageReview (reporting on MLCommons MLPerf Storage v3.0) · Independent storage trade press / audited industry benchmark · Global
- Publisher
- StorageReview
- Original publication
- 2026-09-01
- Source retrieved
- 2026-10-30
- Event date
- 2026-09-01
What happened
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.
Why it matters
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 and measured results
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.
Limitations and uncertainty
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.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-30; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Research and state AI buyers struggle to compare AI storage vendors on credible, audited evidence.
- Stakeholders
- Research computing director
- HPC architect
- Procurement officer
- Discovery questions
- 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?
- Value hypothesis
- Audited results indicate FlashBlade//EXA can scale checkpoint throughput near-linearly; benefit at smaller scale must be estimated and tested.
- Potential engagement
- A sizing workshop mapping audited MLPerf configurations to the customer's planned scale and workloads.
- Unsupported claims
- That results transfer to small configurations
- That absent competitors are slower
- End-to-end application speedups
Pre-sales engineering
Role takeaway
- Fit
- Relevant for large training/checkpoint and KV-cache workloads at research-centre scale.
- Architecture and integration
- 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.
- Proof of value
- Checkpoint write time at purchased scale
- KV-cache read throughput under expected concurrency
- Network validation against benchmarked design
Delivery
Role takeaway
- 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.
- Skills and adoption
- Staff need benchmarking and RDMA networking skills.
- Governance checkpoints
- Acceptance test plan approval
- Configuration deviation log
- Results sign-off
- Acceptance criteria
- 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
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
The submitted design split metadata over pNFS/TCP and bulk data over NFSv3 on RDMA; buyers need RDMA-capable networking to approach these figures.
Governance
Who approves, reviews and stays accountable for outcomes?
Supports evaluation criteria that ask for audited benchmark results with disclosed configuration, while noting absent competitors cannot be judged from the round.
Security and privacy
What data, permissions and controls need testing?
No security content.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
No accessibility or workforce findings.
Procurement
What should contracts, pricing and exit terms secure?
Procurements can reference MLPerf Storage v3.0 categories (checkpoint, KV cache) and require submissions at a scale comparable to the proposed configuration.
Operating model
Which teams own the service once it runs?
Encourages separating audited benchmark evidence from deployment evidence in vendor evaluation.
Publication history
- 2026-10-29Everpure · Issue 017 resources
Stable resource ID: storagereview-mlperf-storage-v3-everpure-2026-09-01