NVIDIA Vera Rubin Delivers Up to 2× Training Performance in MLPerf Debut, Signaling a New Phase in AI Infrastructure

Chinese

NVIDIA's Vera Rubin NVL72 platform made its debut in MLPerf Training v6.1 on November 16, 2026, delivering up to twice the per-GPU performance of the company's optimized GB300 NVL72 system on selected workloads. The results also highlight the growing importance of reinforcement learning, large-scale model training and energy-efficient agentic inference as AI infrastructure evolves beyond conventional chatbot applications.

AI-generated concept diagram of AI data center and high-performance computing

A New Architecture Enters the AI Performance Race

NVIDIA has released its latest MLPerf benchmark results, offering an early view of how its Vera Rubin architecture could reshape the economics of large-scale artificial intelligence.

Announced on November 16, the results show that Vera Rubin NVL72 achieved the fastest training times on six of seven workloads in MLPerf Training v6.1.

Across the benchmark suite, NVIDIA platforms recorded the fastest training times, with the company's existing GB300 NVL72 systems retaining the absolute performance record for the largest DeepSeek-V3 training configuration.

The distinction matters. Vera Rubin demonstrated substantial per-GPU efficiency improvements, while GB300 continued to demonstrate the advantages of operating at extremely large cluster scale.

For AI developers, both dimensions are important: how much useful computation an individual accelerator can deliver and how effectively thousands of accelerators can cooperate.

DeepSeek-V3 Benchmark Reveals the Largest Generational Gain

One of the most notable results came from the DeepSeek-V3 671-billion-parameter training benchmark.

Using 1,024 GPUs, Vera Rubin NVL72 completed the workload in approximately 4.14 minutes, compared with 8.03 minutes for GB300 NVL72 at the same GPU count.

That represents nearly a twofold improvement in performance per GPU.

The advantage is particularly significant because mixture-of-experts models require intensive communication among computing devices.

During training, different expert components must exchange information rapidly. The resulting communication demands can limit performance even when individual processors are powerful.

Vera Rubin's results suggest that improvements in system architecture, interconnects and software optimization are becoming as important as raw computational capability.

NVIDIA also reported approximately 97% scaling efficiency when expanding a DeepSeek-V3 training configuration from 256 to 1,024 GPUs.

Meanwhile, its GB300 NVL72 platform completed the same benchmark in approximately 1.38 minutes using 10,240 GPUs, demonstrating the benefits of much larger clusters.

These are benchmark-specific measurements, not universal claims that every AI training workload will run twice as fast.

Reinforcement Learning Becomes an Infrastructure Priority

A particularly important development in MLPerf Training v6.1 is the introduction of an agentic reinforcement-learning post-training benchmark.

Traditional large language model training has primarily emphasized learning from enormous datasets.

Increasingly, however, developers are building AI systems that must execute tools, interact with environments, evaluate intermediate results and improve through feedback.

These workflows require a combination of inference, simulation and repeated model updates.

NVIDIA reported that Vera Rubin completed the Qwen 3.5 397B GRPO reinforcement-learning benchmark in approximately 10.44 minutes using 1,024 GPUs, compared with 13.78 minutes for GB300 NVL72 at the same scale.

The result illustrates how AI infrastructure requirements are changing.

An AI agent that writes software, performs scientific research or manages a business process may generate and evaluate many intermediate steps before completing a task.

Consequently, infrastructure efficiency can no longer be measured solely by the speed of generating individual tokens.

Developers increasingly need to consider the complete computational cost of accomplishing a reliable task.

From Faster Training to More Efficient AI Agents

NVIDIA also highlighted results from MLPerf Endpoints, a benchmark focused on agentic AI workloads.

For a provisional Kimi K3 agentic workload comparison, Vera Rubin NVL72 delivered more than ten times the throughput per watt of GB300 NVL72 under the stated benchmark conditions.

Unlike the verified MLPerf Training results, this particular Vera Rubin Endpoints result was provisional and pending independent verification.

The comparison also involves specific hardware configurations and response-speed requirements, so it should not be interpreted as a universal tenfold reduction in AI inference costs.

Nevertheless, the metric points toward an increasingly important industry priority.

As AI agents become more capable, they may execute dozens or hundreds of model interactions to complete a single user request.

The commercial value of AI infrastructure will therefore depend not only on how quickly a model responds, but also on how many useful tasks a data center can complete within its power and operating constraints.

AIYXL Analysis: The Next AI Race Is About Useful Work per Unit of Infrastructure

The significance of Vera Rubin's MLPerf debut extends beyond a conventional hardware generation upgrade.

The results illustrate three structural changes in the AI industry.

First, training efficiency remains essential even as the market increasingly focuses on inference. Developers still need to train and refine more capable models, and shorter experimentation cycles can accelerate research.

Second, reinforcement learning and agentic workloads are becoming central design considerations for computing platforms.

Third, energy efficiency and system-level optimization are becoming major competitive variables as AI infrastructure expands.

NVIDIA's position is strengthened not only by accelerator performance but also by its integrated software ecosystem, networking technologies and ability to operate at scale.

However, benchmark leadership does not automatically establish real-world cost leadership.

Customers must still evaluate hardware acquisition costs, power consumption, utilization, software compatibility and application-specific performance.

For the broader AI industry, the direction is clear: the next generation of infrastructure competition will increasingly be measured by how much reliable, useful AI work can be completed with a given amount of computing capacity, electricity and capital.

Sources:

  1. NVIDIA Blog, November 16, 2026 — NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance.

  2. NVIDIA Technical Blog, November 16, 2026 — NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1.

  3. MLCommons — MLPerf Training v6.1 benchmark results.

Editorial note: The 2× figure applies to selected verified training benchmark comparisons. The more than 10× throughput-per-watt figure refers to a specific provisional MLPerf Endpoints comparison and should not be generalized to all inference workloads.