[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"aihot-art-98989":3},{"itemId":4,"vertical":5,"category":6,"source":7,"score":8,"title":9,"summary":10,"analysis":11,"url":12,"coverUrl":13,"direction":13,"marketSignal":13,"publishedAt":14},"98989","ai","产品发布\u002F更新","NVIDIA Technical Blog",79,"在NVIDIA GB300 NVL72上创下MoE预训练世界纪录","NVIDIA在GB300 NVL72上创下MoE预训练世界纪录，训练DeepSeek-V3 671B模型达到每GPU 1648 TFLOPs。","NVIDIA GB300 NVL72在DeepSeek-V3 671B MoE模型预训练中创下世界纪录，每GPU达1648 TFLOPs。\n· MoE模型使每Token计算量下降，通信成为扩展瓶颈，GB300 NVL72通过优化通信解决了这一挑战。\n· 该纪录展示了从GPU、网络到系统软件的全栈协同进步。\n· 结果证明，在千卡规模下，系统级设计比单芯片性能更重要。\n影响\u002F看点：这一纪录凸显了NVIDIA在MoE预训练领域的领先地位，为大规模AI训练树立了新标杆。","https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fsetting-a-world-record-for-moe-pre-training-on-nvidia-gb300-nvl72\u002F",null,"2026-07-21 23:00:00"]