close

MLCommons Releases MLPerf Training v6.1 Results, Spotlighting Recent Advances in AI Training

ⓘ This article is third-party content and does not represent the views of this site. We make no guarantees regarding its accuracy or completeness.

SAN FRANCISCO, Nov. 16, 2026 (GLOBE NEWSWIRE) -- Today, MLCommons® announced new results for the MLPerf® Training v6.1 benchmark suite. The two new benchmarks added in this round, and the submissions received, demonstrate the impact of recent architectural advances in AI training systems and the increasing diversity of the hardware used to train them.

“We’ve reached an exciting new threshold for the MLPerf Training benchmark suite, with a higher level of participation in this round, the representation of both established and emerging hardware offerings, and the increased diversity of configurations and deployments in the results,” said Pavan Yalamanchili, MLPerf Training working group co-chair. “Having much more of the matrix of possible systems represented in the published results gives us a clearer view of how they stack up against each other in the most important real-world scenarios for AI training.”

MLPerf Training v6.1 adds two new benchmarks

The MLPerf Training benchmark suite comprises full system tests that stress models, software, and hardware for a range of machine learning (ML) applications. The open-source and peer-reviewed benchmark suite provides a level playing field for competition, driving innovation, performance, and energy efficiency across the industry. The suite's benchmark collection is curated by a panel of experts from the AI community.

Version 6.1 adds two new benchmarks: DLRMv4-HSTU, a recommender; and Qwen3.5-397B, a large-scale LLM post-training workload.

DLRMv4-HSTU is a recommendation benchmark, updating and replacing the DLRM-dcnv2 benchmark included in earlier versions of the MLPerf benchmark. It utilizes the HSTU attention-based architecture and trains on the Yambda dataset, one of the largest open-sourced recommendation datasets available.

The Qwen3.5-397B benchmark is a first-of-its-kind workload for MLPerf Training: large-scale post-training using Group Relative Policy Optimization (GRPO). Much of the recent innovations in AI, such as agentic functionality, rely on post-training activities; in response, the MLPerf community designed and built a new benchmark to test the performance of such systems. Through post-training this benchmark learns to reason on its own to solve software engineering tasks by training on specific tasks such as tool-calling.

“We are constantly striving to ensure that the MLPerf Training benchmark suite is tracking innovations as they become standard practice in the field,” said Shriya Rishab, MLPerf Training working group co-chair. “Post-training has added a new level of complexity to the architecture of AI training, and we’re proud of the work we’ve done over the past two years on this brand-new benchmark test that will deliver important insights on the performance of state-of-the-art AI training systems. And the v6.1 results show that the community is quickly embracing it.”

Increasing diversity of submissions highlights wide variety of training hardware, architectures, and scale

MLPerf Training v6.1 further increased the number of submitters, submitted systems, results, and the scale of typical systems, beyond some of the records set six months ago in v6.0. Participants in this round submitted 153 systems, yielding 250 test results. 57 of those systems, or 37% of the total, were cloud-based, an increase from 29% in v6.0. Twelve systems used 1,000 or more accelerators, with the largest system using 10,240 accelerators. Systems utilized NVIDIA, AMD and Google accelerator hardware, including for the first time NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood).

“As the task domains and accompanying training workloads specialize, we are seeing a shift in the underlying hardware configurations that are being deployed to host that work,” said Yalamanchili. “The MLPerf Training benchmark suite provides essential guidance for those deploying AI training systems, with data on diverse systems across workloads that help organizations identify the ‘sweet spot’ for their own buildout, whether it’s a single-node system or a multi-node cluster with thousands of accelerators.”

Record industry participation points to broad community support for independent, transparent benchmarking standards

The MLPerf Training v6.1 round includes performance results from 24 submitting organizations: AMD, ASUSTeK, Microsoft Azure, Cisco, CoreWeave, Crusoe, Dell, Firmus, Fujitsu, Giga Computing, Google, HPE, Inventec, KRAI, Lambda, MiTAC, Nebius, NVIDIA, Oracle, Quanta Cloud Technology, Supermicro, tiny corp, TTA, University of Florida. “We would especially like to welcome first-time MLPerf Training submitter Crusoe,” said David Kanter, Head of MLPerf at MLCommons. “The fact that so many organizations continue to submit to the MLPerf Training benchmark while new submitters are still joining in shows the value to the community of having reliable benchmark data for new workloads and new systems.”

Robust participation by a broad set of industry stakeholders strengthens the AI ecosystem as a whole and helps to ensure that the benchmark is serving the community’s needs. We invite submitters and other stakeholders to join the MLPerf Training working group and help us continue to evolve the benchmark.

View the results

Please visit the Training benchmark page to view the full results for MLPerf Training v6.1 and find additional information about the benchmarks. This release includes new visualization tools that allow for richer exploration of the results, including comparing performance across systems and ranking results for similar-scale systems. To learn about each submitter’s results, refer to the supplemental information.

About MLCommons

MLCommons is an open AI engineering consortium whose mission is to standardize AI measurement and enable informed decisions. Since 2018, MLCommons has used collective engineering with industry and academia to build and operate industrial-quality benchmarks and metrics that evaluate AI systems, so that customers can understand the trade-offs and make informed decisions.

For additional information on MLCommons and details on becoming a member, please visit MLCommons.org or email participation@mlcommons.org.


Press Inquiries: contact press@mlcommons.org

Primary Logo

Report this content

If you believe this article contains misleading, harmful, or spam content, please let us know.

Report this article

More News

View More

Recent Quotes

View More
Symbol Price Change (%)
AMZN  258.47
-8.19 (-3.07%)
AAPL  353.39
-2.88 (-0.81%)
AMD  507.51
-26.55 (-4.97%)
BAC  50.28
-0.74 (-1.44%)
GOOG  284.39
-8.13 (-2.78%)
META  627.88
-13.81 (-2.15%)
MSFT  523.24
-0.64 (-0.12%)
NVDA  157.73
-5.12 (-3.14%)
ORCL  119.79
-5.56 (-4.44%)
TSLA  379.37
-10.85 (-2.78%)
Stock Quote API & Stock News API supplied by www.cloudquote.io
Quotes delayed at least 20 minutes.
By accessing this page, you agree to the Privacy Policy and Terms Of Service.

Starting at /week.