TLDR: MLCommons has released the results for MLPerf Training v5.1, a key benchmark suite for machine learning training performance. The latest results highlight substantial advancements in generative AI scenarios, driven by increased participation, diverse hardware and software, and improved system scalability.
SAN FRANCISCO – November 12, 2025 – MLCommons®, a leading engineering consortium, today announced the publication of the MLPerf® Training v5.1 benchmark results. These new findings underscore the rapid evolution and increasing sophistication of the artificial intelligence ecosystem, revealing significant performance enhancements across a range of machine learning applications. The MLPerf Training benchmark suite is designed to provide comprehensive, full-system tests that rigorously evaluate models, software, and hardware, establishing a transparent and level playing field for innovation in the AI industry.
This latest round of benchmarks, published on November 12, 2025, showcased remarkable diversity and scale. A record 65 unique systems were submitted by participants, incorporating 12 distinct hardware accelerators and a variety of software frameworks. Notably, nearly half of these submissions were multi-node configurations, representing an impressive 86% increase compared to the v4.1 round conducted a year prior. These multi-node systems utilized various network architectures, often featuring custom solutions to optimize performance.
Also Read:
- Meta’s SPICE Framework: A New Era of Self-Improving AI Systems
- Cerebras Systems Unveils ‘Cerebras for Nations’ to Empower Global Sovereign AI Development
One of the most compelling aspects of the v5.1 results is the substantial performance improvements observed in two benchmark tests specifically focused on generative AI scenarios. These gains have reportedly outpaced the rate of improvement predicted by Moore’s Law, signaling a rapid acceleration in AI training capabilities. Shriya Rishab, co-chair of the MLPerf Training working group, emphasized the importance of these benchmarks, stating, “More choices of hardware systems allow customers to compare systems on state-of-the-art MLPerf benchmarks and make informed buying decisions.” The MLPerf Training working group is dedicated to creating open-source, peer-reviewed benchmarks that drive innovation, performance, and energy efficiency across the entire industry, ensuring fairness and accuracy through a balanced perspective from diverse AI hardware and software experts.
