NVIDIA Exemplar Cloud: Lessons for Unlocking Full Performance on AI Infrastructure

Two AI computing clusters built from identical NVIDIA H100, GB200 NVL72, or GB300 NVL72 systems can deliver materially different training throughput. NVIDIA routinely sees 8% to 12% gaps between partner deployments and the corresponding NVIDIA reference architecture on the same workload, same model, and same global batch size. The cause is often a stack of configuration choices in the kernel and system settings that affect performance. This highlights the importance of careful configuration and optimization to unlock full performance on AI infrastructure.