NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut
nvidia ·
NVIDIA has announced that its Vera Rubin NVL72 system achieved leading performance in its debut submission to MLPerf Inference v6.1. The new platform delivers up to 3.7 times better throughput than the previous GB300 NVL72 on the Qwen3-VL benchmark and up to 2.5 times higher throughput on DeepSeek-R1. These gains are attributed to full-stack codesign, including enhanced Tensor Cores, Transformer Engines, and NVFP4 precision which reduces memory footprint. Additionally, the GB300 NVL72 demonstrated exceptional scaling efficiency, achieving 99% scaling when expanding from one to four racks for DeepSeek-R1 workloads. Software optimizations within the MLPerf suite contributed significantly, with v6.1 results showing up to 1.6x higher performance compared to version 6.0. NVIDIA also highlighted strong results in agentic inference benchmarks, where Vera Rubin showed a 30x improvement over GB300. The broader ecosystem participation included 19 partners demonstrating high performance across various NVIDIA platforms.
nvidia