NVIDIA's Vera Rubin NVL72 platform completed all MLPerf Training v6.1 benchmarks in its first round, covering large language models, image generation, recommender systems, and reinforcement learning post-training. NVIDIA reported up to 2x training performance per GPU compared to GB300 NVL72, and on DeepSeek-V3 the system scaled from 256 to 1,024 GPUs at roughly 97% efficiency, cutting training time from about 16 minutes to about four. In the inaugural MLPerf Endpoints agentic benchmark using Kimi K3, Vera Rubin NVL72 achieved more than 10x the throughput per watt of GB300 NVL72 at equivalent response speeds. NVIDIA also cited leadership on SemiAnalysis AgentX and Artificial Analysis AgentPerf independent benchmarks.
CoreWeave said it was the first cloud provider to submit multi-rack Vera Rubin results in this round. Its DeepSeek-V3 671B result of 17.65 minutes on 256 Rubin GPUs was 2.7x faster than its prior GB200 NVL72 result, and Llama 3.1 8B trained in 7.16 minutes on a single Vera Rubin NVL72 rack, which CoreWeave described as the fastest 72-GPU submission in the round. CoreWeave said all results came from production infrastructure available to customers.
Microsoft Azure is previewing virtual machines powered by Vera Rubin NVL72 for reasoning, agentic, and large-scale training workloads. In MLPerf Training v6.1, Azure scaled to 10,240 NVIDIA GPUs with GB200 NVL72 systems on DeepSeek-V3, which NVIDIA described as the largest scale in the round. Azure's GB300 NVL72 systems led cloud submissions on DeepSeek-R1 interactive throughput by over 10% in MLPerf Inference v6.1, per Signal65. NVIDIA and Microsoft are collaborating on extending confidential computing to bare-metal and rack-scale AI infrastructure and on power efficiency using NVIDIA MaxLPS.