Mlperf Storage V2 0 Results

MLPerf Storage v2.0 Overview

MLPerf Storage is a standardized benchmark suite developed by MLCommons, designed specifically to evaluate storage system performance under AI workloads. Version 2.0, released in August 2025, introduces several upgrades, including larger dataset sizes (e.g., 1TB+ training data) and AI tasks that better reflect real-world scenarios: pretraining and fine-tuning of GPT-3-style models, as well as generative AI inference such as Stable Diffusion.

This test focuses on key metrics: read throughput, write throughput, read latency, and write latency, simulating high-concurrency I/O operations in a multi-node GPU cluster environment.

Test Result Highlights

• Top-tier performance: NVIDIA DGX SuperPOD paired with DDN EXAScaler storage achieved read throughput up to 45 TB/s and write throughput of 32 TB/s on a 256-GPU training task, leading all submissions.
• Inference optimization: Pure Storage FlashArray delivered latency as low as 50μs and throughput exceeding 20 TB/s in the Stable Diffusion inference benchmark, making it suitable for real-time generation applications.
• Networking innovation: Multiple systems adopted NVMe-oF over RoCEv2 or InfiniBand to achieve end-to-end low latency.

Detailed Result Analysis

MLCommons published over 20 submission results, covering a range of configurations from single machines to supercomputing clusters.

Training Benchmark

• NVIDIA + DDN: Read 45 TB/s, completion time reduced by 25% compared to v1.0.
• NetApp + NVIDIA: Best balance with outstanding cost-performance ratio.
• HPE + VAST Data: Demonstrated excellent scalability in large-scale deployments, supporting 1024 GPUs.

Inference Benchmark

• Pure Storage: Latency optimized by 40%, suitable for Llama model inference.
• IBM Storage Scale: Clear advantages in high-density deployments.

All results were strictly audited to ensure reproducibility and fairness. Charts show that linear scalability becomes a key challenge as the number of nodes increases.

Industry Significance

The MLPerf Storage v2.0 results highlight that storage is no longer a bottleneck in AI infrastructure; high-performance storage is becoming standard for AI training accelerators. Intense competition among vendors is driving technology iterations in NVMe, QDR, and more. Organizations such as LMSYS Org are actively participating to support the open-source benchmark ecosystem.

See the full results on the official page.