NVIDIA secured a complete sweep in MLPerf Training 6.0, achieving the fastest training times across all seven benchmarks and the highest per-accelerator performance. The results, published in June 2026, confirm NVIDIA Blackwell as the only platform to submit across the entire test suite, underscoring its dominance in both dense and Mixture-of-Experts training. The benchmark cycle introduced novel workloads, including the 671B-parameter DeepSeek-V3 and GPT-OSS-20B models. NVIDIA deployed rack-scale GB300 NVL72 and GB200 NVL72 systems to dominate these categories. The GB300 delivered up to 1.6 times faster training than the GB200, driven by expanded memory, higher power ceilings, and NVFP4 precision. At unprecedented scale, NVIDIA coordinated 8,192 Blackwell GPUs to train DeepSeek-V3 in approximately two minutes and Llama 3.1 405B in roughly seven minutes. This throughput relies on fifth-generation NVLink Switches for intra-rack communication and Spectrum-X Ethernet alongside Quantum InfiniBand for inter-rack fabric efficiency. Advanced routing and congestion control mitigate the bursty communication patterns inherent to Mixture-of-Experts training, sustaining near-theoretical bandwidth limits. Performance gains are equally dependent on deep software-hardware co-design. NVIDIA implemented full-iteration CUDA graphs to eliminate CPU-GPU synchronization overhead for dynamic routing models, effectively offloading execution entirely to the accelerator. CuTe DSL kernel fusions combined memory-bound and compute-heavy operations, keeping data localized to registers and delivering significant end-to-end speedups. Additional optimizations include MXFP8 attention computation to accelerate precision without sacrificing accuracy, dynamic pipeline stage balancing to reduce idle cycles, and a dedicated communication overlap scheme that achieves near-total communication-computation overlap. These stack-wide improvements, packaged through Megatron Bridge and optimized in recent NeMo releases, enabled continuous performance uplifts without requiring silicon changes. Production readiness was validated through rigorous resiliency engineering. NVIDIA’s reliability engine continuously monitors chip health, enabling automatic fault routing and self-healing without workload interruption. The network fabric reroutes around failed links in milliseconds, while the NVIDIA Resiliency Extension minimizes recovery time by resuming from recent checkpoints rather than restarting entire jobs. This reliability has attracted widespread industry adoption. Cloud providers and AI developers, including CoreWeave, Google Cloud, and Nebius, are deploying Blackwell fleets to accelerate frontier model research, agentic AI platforms, and large-scale generative media pipelines. The MLPerf Training 6.0 results establish NVIDIA’s full-stack architecture as the industry baseline for scalable AI training. By converging hardware scale, software agility, and enterprise-grade resiliency, NVIDIA compresses previously multi-month training cycles into hours or minutes. This trajectory positions the Blackwell platform to underpin the next generation of frontier models, offering developers measurable efficiency dividends as software optimizations continue to mature.