NVIDIA introduced its Vera Rubin AI platform with a commanding performance in the latest MLPerf benchmarks, released on November 16, 2026. In its debut, the platform swept every MLPerf Training v6.1 benchmark and achieved breakthrough results in its first MLPerf Endpoints submission, showing over tenfold improvement in throughput per watt for agentic AI tasks compared to the previous generation.
MLPerf Training v6.1 measures how quickly systems can train various AI models to a target accuracy, spanning categories like large language models, image generation, recommender systems, and reinforcement learning post-training. Vera Rubin’s NVL72 system outperformed the prior NVIDIA GB300 NVL72 by up to 2x in performance per GPU, demonstrating significant generational improvements in training speed.
Scalability was a highlight: on the DeepSeek-V3 benchmark, Vera Rubin maintained about 97% efficiency when scaling from 256 to 1,024 GPUs, reducing training time from 16 minutes to roughly 4 minutes. NVIDIA also demonstrated operation at a massive 10,240 GPUs scale using GB300 systems, underscoring the platform’s data center readiness for extreme workloads.
These performance gains derive from tight hardware and software integration. NVIDIA NVLink provides high-speed scale-up networking within systems, while Quantum-X800 InfiniBand and Spectrum-X Ethernet enable scale-out networking across data centers. Complementing this is an optimized CUDA-powered software stack that continuously evolves to enhance performance over time.
Agentic inference workloads—which require AI agents to reason, use tools, and verify outputs for complex tasks—benefit notably from Vera Rubin. The MLPerf Endpoints benchmarks, evaluating throughput per watt on real agentic AI scenarios, showed Vera Rubin delivering over 10x the efficiency of earlier NVIDIA systems like the GB300 NVL72, enabling support for far more users under power constraints.
NVIDIA’s achievements extend beyond raw speed to energy efficiency, providing higher throughput per watt and better user concurrency without sacrificing response times. This empowers organizations to deploy smarter, faster AI agents in applications ranging from enterprise workflows to scientific research.
Support from a broad partner ecosystem including ASUS, Microsoft Azure, Google Cloud, Dell, HPE, and others highlights how NVIDIA’s open-source CUDA framework and programmable platform foster innovation across diverse AI workloads. Vera Rubin also demonstrated leadership on independent industry benchmarks, confirming its versatility beyond MLPerf.
While Vera Rubin marks a major advance, the benchmarks reflect idealized, managed testing environments. Real-world deployment will still require careful integration with existing infrastructure, and continuous software optimization will be necessary to sustain and expand performance gains. Nonetheless, Vera Rubin sets a new baseline for large-scale AI training and inference systems.
Source & further reading
NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance ↗Source published 2026-11-16. Claims and limitations should be checked against the original announcement or research.