MLCommons Releases MLPerf Training v6.0 Results

MLCommons released new results for the MLPerf Training v6.0 benchmark suite, citing changes in submission patterns and technical variation. The update added two benchmarks and reported shifts in how training systems were built and deployed, including more cloud participation.

MLPerf Training is described as a benchmark suite of full system tests that stress models, software, and hardware across machine learning applications. The suite is open-source and peer-reviewed, and its benchmark collection is curated by a panel of experts from the AI community.

Version 6.0 added two benchmarks that emphasize sparse computation using a Mixture-of-Experts architecture. DeepSeek V3 is listed as a large-scale pretraining model with 671 billion total parameters and 37 billion activated per token, while GPT-OSS 20B is listed with 21 billion total parameters and 3.6 billion activated per token.

In the v6.0 round, participants submitted 95 unique systems using thirteen different hardware accelerators and 19 different host processors, with 60% of systems described as multi-node. The release also listed performance results from 24 submitting organizations, and it attributed diversity of submissions to multiple FP4-precision recipes used in the results.

“It’s an exciting moment for the community,” said Shriya Rishab, MLPerf Training Working Group co-chair. “We’re seeing strong convergence on a set of best practices for training AI models, but at the same time there is increasing technical diversity in the underlying frameworks and systems that are being used to host and run them.” “There are more ways of getting your AI training than ever before,” said Pavan Yalamanchili, MLPerf Working Group co-chair.

Provided by Globe Newswire on behalf of MLCommons. Click to read original content.