Congratulations to MLPerf Training 6.1 submitters!
David Kanter
Hello Everyone,
Good afternoon! It is my delight to announce MLPerf Training v6.0 on behalf of our remarkable team and entire community of submitters.
The AI ecosystem continues to evolve at a breathtaking pace, and this round captures some of the most significant shifts we've seen yet. We set new records for submission diversity - 95 unique systems from 24 organizations, utilizing 13 different hardware accelerators and 19 different host processors - and introduced two new benchmarks that reflect the industry's decisive turn toward sparse computation and Mixture-of-Experts. This is a real high-water mark for MLPerf Training.
Congratulations to the MLPerf Training working group on the latest iteration of the benchmark suite. The press release is on our website, and you can help promote it on social media:
–LinkedIn: https://www.linkedin.com/feed/update/urn:li:activity:7472667516339683328
–Twitter/X: https://x.com/MLCommons/status/2066902093684998374
–Bluesky: https://bsky.app/profile/mlcommons.org/post/3mog3kul3p22p
Here are some highlights from Training v6.0:
--Two new Mixture-of-Experts (MoE) benchmarks: DeepSeek V3 (671B total parameters, 37B activated per token) and GPT-OSS 20B (21B total parameters, 3.6B activated per token) - reflecting how MoE architectures have become central to the state of the art in large-scale training
--DeepSeek V3 is now the largest benchmark in the suite, pushing the frontier of what MLPerf Training can measure
--GPT-OSS 20B provides an accessible entry point for organizations training at smaller scale, broadening participation in this benchmark
--95 unique systems submitted across 13 different hardware accelerators - a striking demonstration of ecosystem diversity
--60% of results were from multi-node systems, reflecting the industry's continued push toward large-scale distributed training
--Cloud system submissions doubled compared to v5.1, underscoring the growing role of cloud infrastructure in AI training
--MLPerf Training results now show the numerical precision in the results table tooltip
--Multiple different FP4-precision implementations were submitted, showcasing the breadth of hardware and software innovation underway
I want to offer particular congratulations and a warm welcome to our first-time MLPerf Training submitters: Inventec, Netweb Technologies India LTD, TTA and Vultr. Submitting to MLPerf is no small undertaking - your participation raises the bar for the entire community. Your teams and organizations should be proud!
Participants and Organizations this round include: AMD, ASUSTeK, Azure, Cisco, CoreWeave, Dell, Fujitsu, GigaComputing, Google, HPE, Inventec, Krai, Lambda, MITAC, Nebius, Netweb Technologies India LTD, NVIDIA, Oracle, Quanta Cloud Technologies, SCITIX, Supermicro, tinycorp, TTA, and Vultr.
Partners & Working Group Members
As always, MLPerf is a team activity, and this round had a wonderful cast of heroes and heroines:
--Shriya Rishab and Pavan Yalamanchili, who co-chair the Training WG and whose insights into MoE architecture drove the direction of this release. Paul Baumstarck our outgoing co-chair who heavily contributed as well.
--Sarthak Arora, Su-Ann Chong, Ravi Dwivedula, Miro Hodak, Michal Marcinkiewicz, and Shriya Rishab, who led the GPT-OSS 20B taskforce
--Denys Fridman, Michal Marcinkiewicz, Shriya Rishab, Qinwen Xu, and Parmita Mehta, who led the DeepSeek V3 benchmark taskforce
--Alejandro Caruso, who led this round from the MLC side with great skill, and Scott Wasson for guidance
--Greg Diamos for helping with the review process
--Dave Graham and Lori Blonn, who drove the blog and handled our marketing and communications.
--Kevin Schofeld, who wrote and revised the press release
--Pablo Gonzalez Mesa, Karl Pietri, and Joe Kacmarcik for handling all our IT and infrastructure
Explore the full results and technical details at MLCommons: https://mlcommons.org/2026/06/mlperf-training-v6-0-results/
For a deeper look at the new benchmarks, see our supplemental blogs:
DeepSeek V3: https://mlcommons.org/2026/05/deepseek-v3-training-v6-0/
GPT-OSS 20B MoE: https://mlcommons.org/2026/05/gpt-oss-moe-training6/
Again, congrats to everyone - let's keep working to do our part to make AI faster, more capable, and more energy efficient.
Thanks,
David Kanter, Founder
Head of MLPerf
MLCommons