NVIDIA’s Vera Rubin Platform Debuts in MLPerf Benchmarks with Leading Performance Across Training and Agentic Inference
In its first MLPerf benchmarks submission, NVIDIA’s Vera Rubin platform has emerged as the top performer across training, post-training, and agentic inference categories. According to NVIDIA’s official blog post, the platform delivered more than 10x generational gains in throughput per watt for agentic AI — a metric that measures how much agentic work a system can complete per unit of power.

The company claims these results are just the beginning, with ongoing software optimizations expected to further enhance performance over time.
The details
NVIDIA swept every MLPerf Training v6.1 benchmark, covering large language models, image generation, recommender systems, and reinforcement learning post-training. In its debut, the Vera Rubin NVL72 system achieved up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system — a result NVIDIA describes as early evidence of its generation-over-generation innovation.
Taking a break from the news? New players at our sponsor Slotland can get 43 dollars free, no deposit. Sponsored link: Undeniable Progress may earn a commission.
UP Academy: free AI job-skills courses
The platform also demonstrated strong scaling performance. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency, reducing training time from 16 minutes to roughly four.
NVIDIA also ran the benchmark on 10,240 GPUs using GB300 systems — the largest scale in this round — matching Microsoft Azure’s 10,240-GPU setup using GB200 systems.
NVIDIA attributes this performance to its NVLink scale-up networking, which connects GPUs within each system, and Quantum-X800 InfiniBand and Spectrum-X Ethernet for scale-out networking across data centers. Combined with optimized software, these technologies enable thousands of GPUs to work in concert, helping developers build more intelligent models while cutting training times.
NVIDIA’s MLPerf Training v6.1 results were retrieved from www.mlcommons.org on November 16, 2026, from multiple entries including 6.1-0116, 6.1-0030, 6.1-0135, 6.1-0020, 6.1-0049, 6.1-0123, 6.1-0138, 6.1-0073, 6.1-0124, 6.1-0129, and 6.1-0038. The MLPerf name and logo are trademarks of MLCommons Association.
The company’s blog post also notes that AI agents are now integral to business workflows, scientific research, and customer service — and faster training enables these agents to reach users sooner, adapt to new threats, and advance discovery.
[Attribution: NVIDIA, https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin/; MLPerf results retrieved from www.mlcommons.org on Nov 16, 2026.]
