NVIDIA Blackwell Dominates MLPerf Training 6.0 as AI Infrastructure Race Intensifies

NVIDIA's Blackwell platform swept the MLPerf Training 6.0 benchmarks, achieving the fastest training times across all seven categories, including new mixture-of-experts workloads. The results, published June 16, 2026, mark the first time a single platform has led every benchmark in the suite.

On the largest MoE model, DeepSeek-V3 671B, NVIDIA scaled its submission to 8,192 GPUs using GB200 NVL72 systems, the largest Blackwell-based cluster in MLPerf Training history. CoreWeave delivered the fastest time to train for that benchmark, reaching the quality target in 2.02 minutes at the same scale using GB300 NVL72 systems connected with Spectrum-X Ethernet.

Blackwell Ultra delivers up to 1.6x performance gain

The GB300 NVL72 system, powered by Blackwell Ultra, delivered up to 1.6x faster training than the GB200 NVL72 at the same scale. Key improvements include higher compute density with NVFP4 precision, expanded memory capacity, and a higher power ceiling that sustains peak performance. NVIDIA also demonstrated NVFP4 training methods that meet strict accuracy requirements across large and small scale pretraining and fine tuning workloads.

Microsoft Azure scaled Llama 3.1 405B training to 8,192 GPUs using GB200 NVL72 systems, reaching the reference quality target in 7.07 minutes, the fastest time for that benchmark. Nineteen organizations submitted results this round, including ASUSTeK, Cisco, Dell Technologies, Fujitsu, and Google Cloud.

Copilot Cowork goes GA, Siemens adopts agentic workflows

Microsoft and Anthropic announced that Copilot Cowork is now generally available to all users. The tool aims to help workers tackle complex tasks by combining AI assistance with collaborative features. Separately, Siemens is using agentic workflows on Google Cloud to modernize legacy codebases spanning hundreds of millions of lines developed over more than a decade. The approach, which Siemens calls "slicing the elephant," breaks down monolithic code into manageable pieces for AI assisted refactoring.

In the open source community, Subquadratic released SubQ 1.1 Small, a new model architecture that promises subquadratic attention mechanisms for more efficient inference. Meanwhile, developer Vicki Boykis published a widely shared post arguing that running local models has become practical, citing improvements in quantization, hardware support, and model quality.

The MLPerf results underscore the growing importance of infrastructure reliability at scale. NVIDIA's platform includes self healing capabilities that automatically route around detected faults without interrupting workloads, and the NVRx extension minimizes time lost when faults do occur by resuming from recent checkpoints rather than restarting entire jobs.