NVIDIA Exemplar Cloud：在 AI 基础设施上释放完整性能的经验
NVIDIA Exemplar Cloud: Lessons for Unlocking Full Performance on AI Infrastructure
NVIDIA 指出，采用相同 H100、GB200 NVL72 或 GB300 NVL72 的两套 AI 集群，在同一工作负载、模型和全局 batch size 下，训练吞吐量常出现 8% 到 12% 的差距，原因多来自内核层面的配置选择。NVIDIA 通过 Exemplar Cloud 总结在 AI 基础设施上释放完整性能的经验。
Two AI computing clusters built from identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 systems can deliver materially different training throughput. We...
Two AI computing clusters built from identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 systems can deliver materially different training throughput. We routinely see 8% to 12% gaps between partner deployments and the corresponding NVIDIA reference architecture (RA) on the same workload, same model, same global batch size. The cause is often a stack of configuration choices in the kernel…
来源：NVIDIA Developer Blog · developer.nvidia.com