Vera Rubin NVL72 reports training performance and GRPO results in MLPerf Training v6.1
Key details NVIDIA said Vera Rubin NVL72 recorded the shortest training times in six of seven benchmarks in its first participation in MLPerf Training v6.1. For DeepSeek-V3 671B, with both systems using 1,024 GPUs, it reported 4.14 minutes for Vera Rubin and 8.03 minutes for GB300. This evaluation newly included a Qwen 3.5 397B GRPO reinforcement learning post-training benchmark. Use cases Suggested use (inference): When comparing large-scale training infrastructure, results using the same model and GPU count can be considered separately. The source evaluates not only pretraining but also GRPO workloads that repeatedly generate rollouts, score them, and update weights. Costs and limitations The DeepSeek-V3 comparison consists of benchmark results from configurations using 1,024 GPUs each and does not guarantee performance on general workloads. Some figures from reruns with the latest software were explicitly identified as unverified. Usage fees and licensing are unconfirmed.
Visit original source