Fastest, Largest, Most Powerful: NVIDIA Blackwell Sweeps MLPerf Training 6.0

NVIDIA's Blackwell platform swept every category in the latest MLPerf Training 6.0 benchmark, demonstrating overwhelming performance. NVIDIA was the only platform submitted across all benchmark categories, and it recorded the fastest training time in every category.

This round of benchmarks added two new pretraining workloads, DeepSeek-V3 671B and GPT-OSS-20B, reflecting the growing importance of Mixture-of-Experts (MoE) architectures. NVIDIA effectively addressed the communication bottlenecks that arise during large-scale MoE training by leveraging the high bandwidth of NVLink.

Key achievements include the following:

• Achieved the fastest training speed in every benchmark
• Performed large-scale training at a scale of 8,192 GPUs using the NVIDIA Blackwell NVL72 system
• The GB300 NVL72 system delivered up to 1.6x faster performance compared to the equally sized GB200 NVL72

NVIDIA also showcased an innovation with its NVFP4 low-precision training technology, maximizing performance while maintaining accuracy. This was also applied to the recent pretraining of the NVIDIA Nemotron 3 Ultra model.