NVIDIA Blackwell Delivers Fastest Results Across All MLPerf Training 6.0 Tests
HPCwire, Tuesday, June 16th, 2026
NVIDIA Blackwell swept MLPerf Training 6.0, posting the fastest time-to-train across all seven benchmarks.
HPCwire reports that NVIDIA's Blackwell platform was the only one submitted across every MLPerf Training 6.0 benchmark and delivered the fastest time to train on all seven.
Microsoft Azure scaled Llama 3.1 405B training to 8,192 GPUs on GB200 NVL72 systems, hitting the quality target in 7.07 minutes, while CoreWeave set the fastest DeepSeek-V3 671B time at 2.02 minutes using GB300 NVL72 systems with Spectrum-X Ethernet. GB300 NVL72 delivered up to 1.6x faster training than GB200 at the same scale, aided by higher NVFP4 compute density.
The DeepSeek-V3 671B run on 8,192 GPUs was the largest-scale Blackwell submission in MLPerf Training to date.