Skip to main content

Press release · November 16, 2026 · globenewswire.com

MLCommons publishes MLPerf Training v6.1 results

Summary

MLCommons has published results for MLPerf Training v6.1, a round that adds two benchmarks and draws a larger and more varied set of submissions than earlier rounds. The organization describes the results as reflecting recent architectural advances in AI training systems and a growing diversity of the hardware used to train them.

The MLPerf Training suite runs full-system tests that stress models, software, and hardware across a range of machine learning applications. Its benchmark collection is open source, peer reviewed, and curated by a panel of AI community experts, and it is intended to give systems a common basis for comparison while encouraging gains in performance and energy efficiency.

DLRMv4-HSTU replaces DLRM-dcnv2 from earlier versions as the recommendation benchmark, using the HSTU attention-based architecture and training on the Yambda dataset, described as one of the largest open-sourced recommendation datasets available. The Qwen3.5-397B workload is a first-of-its-kind test of large-scale post-training using Group Relative Policy Optimization (GRPO), a stage that underpins much recent agentic functionality. Through tasks such as tool-calling, the model learns to reason independently while working on software engineering problems.

Participation grew across several measures. Submitters sent 153 systems that produced 250 test results from 24 organizations, and cloud-based systems accounted for 57 of those systems, or 37%, up from 29% in v6.0. Twelve systems used 1,000 or more accelerators, with the largest using 10,240. Hardware from NVIDIA, AMD, and Google appeared, including NVIDIA Rubin-generation GPUs and Google TPUv7 (Ironwood) for the first time, and Crusoe made its debut as a submitter.

Pavan Yalamanchili, MLPerf Training working group co-chair, said “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.”

David Kanter, Head of MLPerf at MLCommons, said “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.” The working group invites additional submitters and stakeholders to join and help the benchmark continue to evolve.

Our summary. The original keeps its own byline.

Read the original on globenewswire.com

Companies in this article

Topics