Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72

За 30 секунд

Что произошло

NVIDIA GB300 NVL72 achieved 1,648 TFLOPs per GPU pre-training DeepSeek-V3 671B, a world record for mixture-of-experts model training.

Почему это важно

Matters for engineers scaling large language models who need to understand communication bottlenecks and co-designed hardware-software performance gains.

На что обратить внимание

Results use optimized software stacks; real-world performance depends on framework choice, model architecture, and whether your workload matches DeepSeek-V3 characteristics.