Nvidia has once again demonstrated its commanding leadership in AI training hardware, claiming the top position across all seven workloads in the just-released MLPerf Training v6.0 benchmark suite from MLCommons. The results, published Monday, confirm the Blackwell platform’s status as the gold standard for AI model training — the computationally intensive process that teaches AI systems to recognize patterns and make decisions.
Particularly noteworthy in this round is the introduction of two new benchmarks based on mixture-of-experts (MoE) architectures — DeepSeek-V3 with 671 billion total parameters and GPT-OSS-20B. These benchmarks reflect the cutting-edge architecture that has become dominant in frontier AI development, where sparse computation allows models to activate only a fraction of their parameters for any given input, dramatically improving efficiency without sacrificing capability.
Nvidia submitted results on both its GB200 NVL72 and the newer GB300 NVL72 systems, with the GB300 delivering up to 1.6x faster training compared to its predecessor at the same scale. This generational leap underscores Nvidia’s continued innovation in AI hardware, maintaining the performance improvements that have made it the preferred platform for AI research and production workloads worldwide.
CoreWeave, a cloud provider specializing in GPU infrastructure, posted the fastest DeepSeek-V3 training time in the benchmark, completing the workload in just 2.02 minutes using 8,192 GB300 NVL72 GPUs across 2,048 nodes — the largest GB300 cluster submitted in this round. The result, achieved on CoreWeave’s production cloud infrastructure, demonstrates the commercial viability of training frontier models in remarkably compressed timeframes.
The record participation in MLPerf v6.0 itself tells a compelling story about AI infrastructure growth. The round attracted 24 submitting organizations using 95 unique systems across 13 different hardware accelerators, with cloud submissions more than doubling compared to six months ago. This reflects the explosive expansion of AI training infrastructure globally.
Nvidia’s cumulative MLPerf training and inference wins since 2018 now stand at more than 290 — nine times the total of all other submitters combined. This unparalleled track record of performance leadership has made Nvidia the foundational infrastructure layer for the AI revolution, with its CUDA ecosystem, specialized networking, and cutting-edge GPU hardware forming the backbone of AI development from academic research to large-scale commercial deployment.
