Portuguese
MLCommons has officially released the MLPerf Storage v3.0 results. This major iteration not only updates benchmark specifications but also reshapes the competitive landscape of AI storage performance. As the industry’s only audited AI storage benchmark, v3.0 for the first time covers the full AI workflow pipeline, including training throughput, model checkpointing, and two newly added inference workloads: vector database querying and KV cache processing. A total of 19 organizations submitted 143 test results, among which 11 are first-time participants, including Microsoft Azure, NVIDIA, Everpure, and a wave of emerging AI storage vendors. Notably, leading vendors from the v2.0 leaderboard—DDN, Huawei, Hammerspace, and Lightbits—did not participate in this round.
This benchmark upgrades simulated accelerators from H100 to B200, raising the bar for per-accelerator bandwidth performance. As a result, v3.0 results are not comparable with v2.0 data. All listed outcomes are vendor-submitted results audited by MLCommons, without independent verification or retesting by StorageReview.
Checkpointing Stands Out with Breakthrough Throughput Performance
MLCommons upgrades checkpointing to a first-class workload, as it has become a critical bottleneck for frontier-scale AI training. Large-scale training jobs continuously write massive model state data at fixed intervals, and prolonged checkpoint flushing directly causes accelerator idling and degrades overall training efficiency. The highest throughput metrics in this round are generated from checkpointing tests.
Everpure (formerly Pure Storage), a first-time MLPerf participant, delivers nearly linear performance scalability with its FlashBlade//EXA platform. Its checkpoint write bandwidth reaches 327.6 GiB/s at 10 data nodes, 484.1 GiB/s at 15 nodes, and 655.6 GiB/s at 20 nodes. The 30-node configuration achieves a record-breaking 877.5 GiB/s write throughput and 833.0 GiB/s read throughput. This marks the first verified, audited benchmark evidence validating Everpure’s long-claimed high-throughput performance capabilities.
Another industry milestone comes from Azure Managed Lustre, the first hyperscale cloud storage service to join the benchmark. Deployed with 4,096 TiB of capacity and 128 clients, it delivers 642.2 GiB/s checkpoint write throughput, enabling cloud-native storage to compete with purpose-built AI storage arrays in top-tier performance.
Top Checkpointing Submissions
|
System
|
Write Bandwidth (GiB/s)
|
|---|---|
|
Everpure FlashBlade//EXA (30 data nodes)
|
877.5
|
|
Azure Managed Lustre (4,096TiB, 128 clients)
|
642.2
|
|
YanRongTech F9000X (8 clients)
|
313.7
|
|
Suzhou Zishan Longlin (multiple configurations)
|
313.7
|
|
TuringData F9200 (8 clients)
|
307.1
|
|
UBIX UbiPower18000 (3 storage nodes)
|
293.8
|
Training Leaderboard Dominated by Emerging Vendors
The 3D U-Net training workload leaderboard is occupied by new industry entrants. YanRongTech’s F9000X tops the list with a sustained read throughput of 543.9 GiB/s, supporting 99 simulated B200 accelerators and securing the highest training throughput in this round. Singapore-based AI infrastructure startup TuringData makes its benchmark debut with 541.5 GiB/s, powering 96 simulated B200s via only three storage nodes. Its three-node F9200 cluster also achieves 539.9 GiB/s read and 307.1 GiB/s write throughput in 70B-scale checkpointing tests. UBIX’s UbiPower18000 delivers 454.1 GiB/s throughput with three storage nodes equipped with 16 × 15.36TB NVMe drives and quad NDR4000 networking. HPE is the only established enterprise storage vendor with valid training results, scoring 345.8 GiB/s with its K3000 platform.
Top 3D U-Net Training Submissions
|
System
|
Read Bandwidth (GiB/s)
|
Simulated B200 Accelerators
|
|---|---|---|
|
YanRongTech F9000X
|
543.9
|
99
|
|
TuringData F9200
|
541.5
|
96
|
|
UBIX UbiPower18000
|
454.1
|
80
|
|
Suzhou Zishan Longlin
|
433.4
|
80
|
|
FarmGPU
|
406.7
|
75
|
|
Azure Managed Lustre
|
379.1
|
70
|
Inference Workloads Officially Integrated into Benchmark
The two newly added inference workloads reflect the rapid growth of real-world AI storage demands. The KV cache test evaluates storage performance for paging attention states in and out of memory during long-context LLM inference, a critical scenario analyzed in-depth in long-context inference research. Everpure again takes the lead with 1,623.4 GiB/s KV cache read throughput on a 51-host FlashBlade//EXA deployment. UBIX (443.7 GiB/s) and FarmGPU (443.6 GiB/s) lead standard-scale KV cache submissions. For vector database workloads, TTA’s Seahorse system achieves the top query throughput of 57,620 queries per second.
v3.0 also adds native S3 object storage support, with approximately one-sixth of all submissions leveraging the protocol. NVIDIA dominates object storage entries with 20 AIStore submissions across OCI, AWS, and GCP environments, peaking at 133.3 GiB/s checkpoint write throughput. Though trailing parallel file systems in raw speed, object storage’s successful adaptation to training and checkpointing workloads marks a major shift from previous benchmarks, where POSIX file systems monopolized high-performance AI workloads. Additionally, MLCommons introduces power efficiency and rack space utilization as official core metrics in this round.
Everpure FlashBlade//EXA Performance Breakdown
Everpure’s FlashBlade//EXA leads checkpointing for both 405B and 1.25T parameter models, as well as KV cache retrieval benchmarks, demonstrating outstanding sustained throughput for data-heavy AI training and inference workloads.
The tested configuration consists of 30 FlashBlade//EXA blades (120 DirectFlash Modules for metadata processing) and 30 Linux/NVMe data nodes, delivering approximately 866 TB of usable capacity within a single unified file system. Metadata is transmitted via pNFS/TCP, while bulk data transfers use NFSv3 over RDMA. In the 1.25T parameter checkpointing simulation with 1,024 client accelerators, the 30-node setup sustains 877.52 GiB/s write bandwidth over 17.74 seconds and 588.28 GiB/s read bandwidth in 28.99 seconds. In KV cache inference tests, it achieves 85,736 tokens/sec (Llama 3.1 8B, storage-only), 67,642 tokens/sec (8B model, storage + system memory), and 33,403 tokens/sec (70B, storage-only).
Industry Implications of Absent Vendors
Benchmark results reflect both submitters’ capabilities and absent vendors’ strategic choices. Top vendors dominating large-scale neocloud and AI factory deployments—VAST Data, WEKA, and DDN—alongside former v2.0 leaders Huawei, Hammerspace, and Lightbits, opted out of v3.0 submissions. Vendor no-shows typically stem from engineering cycle arrangements, product release timelines, or benchmark strategy adjustments, and do not equate to inferior performance.
However, this creates a clear divergence: the audited benchmark leaderboard no longer aligns with real-world large-scale deployment leaders. Industry buyers must evaluate both audited benchmark data and practical deployment track records. StorageReview’s updated 2026 enterprise storage array ranking page fully documents this discrepancy with complete v3.0 result data.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!