NVIDIA publishes Vera Rubin NVL72 MLPerf Endpoints v1.0 results

developer.nvidia.com5 min read

Nvidia has published provisional MLPerf Endpoints v1.0 benchmark results showing that its Vera Rubin NVL72 platform achieves over 10x the throughput per kilowatt of provisioned power compared to the GB300 NVL72. The benchmark evaluates agentic AI workloads using coding and enterprise workflows. According to the company, the performance gains are driven by hardware advancements like HBM4 memory alongside software optimizations including prefix caching, disaggregated serving, and model-specific execution kernels. Open article →