Blackwell and GB200 NVL72 debut in MLPerf Training v5.0 with up to 2.6x per-GPU gains over Hopper

NVIDIA/Blackwell

NVIDIA was the only company to submit results on every MLPerf Training v5.0 benchmark, reporting 2.2x performance on Llama 3.1 405B pretraining and 2.5x on Llama 2 70B fine-tuning with eight Blackwell GPUs versus prior-generation Hopper systems.

• Blackwell submitted across all v5.0 benchmarks; NVIDIA only vendor with complete coverage and highest performance at scale.
• 2.2x performance on Llama 3.1 405B pretraining (GB200 NVL72); 2.5x on Llama 2 70B LoRA fine-tuning (DGX B200) vs. Hopper baseline.
• Gains reflect advances in architecture, memory bandwidth, interconnect and software frameworks like NVIDIA NeMo.