NVIDIA Blackwell tops all MLPerf training benchmarks

NVIDIA announced its Blackwell platform achieved the fastest training results in all 7 benchmarks of MLPerf Training 6.0, making it the only platform to cover every test. The benchmarks include major AI models like DeepSeek-V3, GPT-OSS 20B, and Llama. NVIDIA submitted two rack-scale systems, GB200 NVL72 and GB300 NVL72, each linking 72 GPUs with 5th-gen NVLink Switch. The GB300 NVL72 delivers up to 1.6x faster training than the GB200 NVL72. NVIDIA also scaled the Blackwell cluster to 8,192 GPUs, setting the largest Blackwell record in MLPerf Training for the DeepSeek-V3 task. CoreWeave used the GB300 NVL72 with Spectrum-X Ethernet to train DeepSeek-V3 in just 2.02 minutes at the same scale.