GB200 NVL72
GB200 NVL72是NVIDIA推出的一款高密度AI计算系统，采用Blackwell架构，将多颗GB200超级芯片（GPU与CPU组合）集成于单一机架单元中。该系统通过NVLink高速互联技术将各计算节点紧密连接，形成统一的大规模并行计算平台，主要面向大型语言模型训练、推理及高性能科学计算等密集型AI工作负载场景。
Timeline (last 90 days)
Prime Intellect launched Prime Inference on October 2, 2025, and achieved production data showing the same number of GPUs carrying 75% more requests on GB200 NVL72.
vLLM provides day-0 support for NVIDIA Vera Rubin NVL72, reporting 7.8x throughput per GPU versus GB200 NVL72 (AgentX) and 3.7x VLM throughput versus GB300 (MLPerf v6.1), with locality domains tuned for MoE routing.
The NVIDIA GB200 NVL72 adopts a liquid-cooled rack with approximately 130 TB/s communication bandwidth, presenting extremely high system engineering barriers.
NVIDIA has launched rack-level systems such as the GB200 NVL72.
All Facts (15)
NVLink Switch的拓扑并非完全扁平，经过不同层级Switch的通信路径在延迟和有效带宽上存在差异
90%UnverifiedSlurm支持拓扑感知作业调度功能，可用于优化GB200 NVL72系统的工作负载放置
90%UnverifiedGB200 NVL72在FP8精度下的AI算力可达1.44 ExaFLOPS
90%Unverified整个NVL72系统的总显存容量超过13TB
90%Unverified传统Slurm调度器以节点为基本单位分配资源，在NVL72架构下会导致资源碎片化和通信效率下降问题
85%UnverifiedSlurm的topology plugin可以将NVL72的NVLink拓扑信息注入调度器，实现拓扑感知的分配决策
85%UnverifiedGB200 NVL72的72块GPU可以按照NVLink拓扑结构划分为不同粒度的块，包括9个8-GPU块、4个18-GPU块、2个36-GPU块或1个72-GPU块
85%UnverifiedGB200 NVL72是面向超大规模AI训练和推理场景的旗舰级计算平台
85%UnverifiedGB200 NVL72将NVLink互联从单节点内部扩展到了整个机架级别，机架内任意两块GPU之间都能通过NVLink直接通信
70%UnverifiedGB200 NVL72中的72颗GPU通过NVLink Switch芯片实现全互联
70%UnverifiedNVIDIA GB200 NVL72在单个机架内集成了72颗Blackwell GPU
70%UnverifiedThe NVIDIA GB200 NVL72 adopts a liquid-cooled rack with approximately 130 TB/s communication bandwidth, presenting extremely high system engineering barriers.
50%UnverifiedNVIDIA has launched rack-level systems such as the GB200 NVL72.
50%UnverifiedPrime Intellect launched Prime Inference on October 2, 2025, and achieved production data showing the same number of GPUs carrying 75% more requests on GB200 NVL72.
50%Unverified随着AI模型参数量持续膨胀，单个训练作业跨越多个NVL72机架将成为常态
75%