Congratulations to MLPerf Storage - v3.0 results are live!!!!
David Kanter
Hello Everyone,
Good morning! It is my delight to announce MLPerf Storage v3.0 on behalf of our remarkable team and entire community of submitters.
MLPerf Storage measures storage system performance for AI and ML workloads, giving the industry a fair, reproducible way to compare how well different storage architectures keep up with training, checkpointing, and now GenAI serving infrastructure. This release marks our first submission round under the required-membership model and expands the benchmark's scope into new AI-serving storage patterns and object storage solutions.
Congratulations to the MLPerf Storage working group on the latest iteration of the benchmark suite. The blog is on here, and you can help promote it on social media:
Here are some highlights from Storage v3.0:
First submission round under the required-membership model — 19 organizations submitted, including 11 first-time submitters, showing tremendous community engagement.
Expanded benchmarks that cover AI Inference: the new KV cache and vector database complement training and checkpointing workloads and address the fastest growing AI workloads
Object storage support to enable broader participation with adoption in 16% of results
Continued emphasis on reproducibility, standardized evaluation, and normalization across a broad range of storage architectures and vendors.
I want to offer particular congratulations and a warm welcome to our first-time MLPerf Storage submitters: Azure, Everpure, HolmesAI, NVIDIA, Nebius, NewFW, OpenLake, Suzhou Zishan Longlin, TuringData, XSKY, ZettaLane. Submitting to MLPerf is no small undertaking — your participation raises the bar for the entire community.
Participants and Organizations: Alluxio, Azure, Everpure, FarmGPU, HPE, HolmesAI, InspurData, NVIDIA, Nebius, NewFW, OpenLake, SAMSUNG, Suzhou Zishan Longlin, TTA, TuringData, UBIX, XSKY, YanRongTech, ZettaLane
Partners & Working Group Members: As always, MLPerf is a team activity, and this round had a wonderful cast of heroes and heroines:
Curtis Anderson, Brian Belgodere, Huihuo Zheng and Oana Balmau who chaired the Storage Working Group and guided this release.
Russ Fellows, Devasena Inupakutika, David Slik and Hazem Awadallah who worked on key portions of the benchmark rules, infrastructure, and workloads and provided invaluable feedback
Alejandro Caruso who ably shepherded this release to the finish line and helped build out an extensive set of charts for our results
Tanvi Gour, Elizabeth Walker, Pablo Gonzalez Mesa, Karl Pietri, and Joe Kacmarcik who led this round from the MLC side and for handling all our IT and engineering infrastructure.
Lori Blonn, who handled our marketing and communications.
Kevin Schofield, who wrote and revised the blog article.
Explore the full results and technical details at MLCommons:
Results
Press Release
Supplemental
Again, congratulations to everyone - let's keep working together to make AI infrastructure faster, more efficient, and more accessible.
Thanks,
David Kanter
Founder, Head of MLPerf
MLCommons