NVIDIA Blackwell GB300 Systems Dominate MLPerf 6.0 Benchmarks

NVIDIA Blackwell GB300 systems have secured the top spot across every benchmark in the latest MLPerf Training 6.0 suite. The results establish a clear performance lead over competing hardware from AMD and other vendors. Microsoft Azure demonstrated practical scale by training Llama 3.1 at 405 billion parameters using 8,192 GB200 GPUs.

GB300 configurations deliver up to 60 percent faster training speeds than previous generations across all seven tests.

The testing platform relies on the Blackwell architecture running in dual-chip NVL72 configurations. NVIDIA reports that these upgraded GB300 setups deliver up to 60 percent faster training speeds compared to previous GB200 generations. The company utilized NVFP4 precision modes to maximize AI compute density across its data centers.

CoreWeave achieved the quickest completion time for DeepSeek-V3 at 8,192 GPU scale with a finish of just over two minutes. Microsoft Azure reached quality targets for Llama 3.1 training in approximately seven minutes using similar system configurations. NVIDIA served as the sole vendor to submit results across all seven required tests in the benchmark suite.

Competitors like AMD declined to participate in several specific evaluations including DeepSeek-v3 and GPT-OSS 20B. The absence of rival submissions highlights a significant gap in available data for direct comparison against Blackwell systems. These benchmarks focus exclusively on large language model training workloads rather than inference tasks.